title
stringlengths 1
300
| score
int64 0
8.54k
| selftext
stringlengths 0
40k
| created
timestamp[ns]date 2023-04-01 04:30:41
2025-06-30 03:16:29
⌀ | url
stringlengths 0
878
| author
stringlengths 3
20
| domain
stringlengths 0
82
| edited
timestamp[ns]date 1970-01-01 00:00:00
2025-06-26 17:30:18
| gilded
int64 0
2
| gildings
stringclasses 7
values | id
stringlengths 7
7
| locked
bool 2
classes | media
stringlengths 646
1.8k
⌀ | name
stringlengths 10
10
| permalink
stringlengths 33
82
| spoiler
bool 2
classes | stickied
bool 2
classes | thumbnail
stringlengths 4
213
| ups
int64 0
8.54k
| preview
stringlengths 301
5.01k
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Llama 4 smells bad
| 0 |
Here is the story of Llama 4 so far, including the LM Arena drama.
| 2025-04-17T21:43:03 |
https://fastml.com/llama-4-smells-bad/
|
Foxtr0t
|
fastml.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1okri
| false | null |
t3_1k1okri
|
/r/LocalLLaMA/comments/1k1okri/llama_4_smells_bad/
| false | false |
default
| 0 | null |
Multi node/ cluster here at home
| 2 |
Want to build a multi-node cluster to play with some of the extensibilities across multiple gpus and I want this cluster to be networked together, not some of the local physically co-located high speed interfaces that exist. Curious if anyone has this kind of hardware setup in their house and maybe some tips or tutorials that they've looked at in terms of the hardware and software stack.
| 2025-04-17T21:46:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1onc2/multi_node_cluster_here_at_home/
|
onemoreburrito
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1onc2
| false | null |
t3_1k1onc2
|
/r/LocalLLaMA/comments/1k1onc2/multi_node_cluster_here_at_home/
| false | false |
self
| 2 | null |
Evaluating a LLM API that conforms to OpenAI specification
| 1 |
I’m looking for a library that will hit an API that has the same format as an OpenAI API but is locally hosted or hosted on a server on my same network.
I’m looking at either LightEval or Lm-evaluation-harness, but it doesn’t seem like either support this, especially if the server doesn’t have an equivalent model pushed to HuggingFace.
Open to any suggestions.
| 2025-04-17T21:50:05 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1oqal/evaluating_a_llm_api_that_conforms_to_openai/
|
larenspear
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1oqal
| false | null |
t3_1k1oqal
|
/r/LocalLLaMA/comments/1k1oqal/evaluating_a_llm_api_that_conforms_to_openai/
| false | false |
self
| 1 | null |
Voice AI Assistant
| 0 |
Trying to set up a voice assistant I can fine tune eventually, but I don’t know where I keep getting it wrong. I’m vibe coding (to be quite fair), using a Jabra 710 as the I/O device. Explored whisper, coqui, but even when I got it to work with the wake word, respond, albeit hallucinating a lot, trying to switch the assistant’s voice is where I got stuck.
It’s not working seamlessly, so getting to the next point of fine-tuning is not even a stage I am at yet. I am using phi-2.
Anyone have a repo I can leverage or any tips on a flow that works. I’ll appreciate it
| 2025-04-17T21:52:46 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1osf2/voice_ai_assistant/
|
SolidRemote8316
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1osf2
| false | null |
t3_1k1osf2
|
/r/LocalLLaMA/comments/1k1osf2/voice_ai_assistant/
| false | false |
self
| 0 | null |
What's the smallest model you've used that has decent success with basic Agents and Tool-Calling ?
| 4 |
Just a few very simple SmolAgents functions right now.
I've noticed that
- Qwen 14B instruct models work well until you quantize them under Q4.
- Phi4 14B can adhere to instructions very well and calls the tools well, but the code logic and args it passes is sometimes wonky.
- Qwen-Coder 14b is very good at calling tools, but there is a creative/reasoning portion to this task that it's poor at
Anything smaller that's worked for you?
| 2025-04-17T21:54:20 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1otp8/whats_the_smallest_model_youve_used_that_has/
|
ForsookComparison
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1otp8
| false | null |
t3_1k1otp8
|
/r/LocalLLaMA/comments/1k1otp8/whats_the_smallest_model_youve_used_that_has/
| false | false |
self
| 4 | null |
Judging Embeddings
| 1 |
[removed]
| 2025-04-17T22:14:41 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1pa2y/judging_embeddings/
|
remyxai
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1pa2y
| false | null |
t3_1k1pa2y
|
/r/LocalLLaMA/comments/1k1pa2y/judging_embeddings/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'VQKIVyFaXVqS_-qOiQM9Mb5NKxNJ2ZmI_eeHhx4p3JA', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=108&crop=smart&auto=webp&s=55d30c9b3fb071b39044614c03cc1ecba14a48c7', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=216&crop=smart&auto=webp&s=fe482a54df69d9741c50bbcb894321457de762e0', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=320&crop=smart&auto=webp&s=837752a5c85c4963f7da3c75bf760afd14ac97b8', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=640&crop=smart&auto=webp&s=d7e456d129a7e0315ba3b18a9709904578fc92c4', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=960&crop=smart&auto=webp&s=310f85edae9aa2bbc19c1374c06bb4751461876e', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=1080&crop=smart&auto=webp&s=7cd9d1a856d474743894c143d5e6d2152af760fa', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?auto=webp&s=9e4243400d6ee75b7bef46e9ca91421f9b9c7393', 'width': 1200}, 'variants': {}}]}
|
$1/Week — Compare ChatGPT, DeepSeek, Gemini & 60+ Models
| 1 |
[removed]
| 2025-04-17T22:16:18 |
[deleted]
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1pbbz
| false | null |
t3_1k1pbbz
|
/r/LocalLLaMA/comments/1k1pbbz/1week_compare_chatgpt_deepseek_gemini_60_models/
| false | false |
default
| 1 | null |
||
$1/Week — Compare ChatGPT, DeepSeek, Gemini & 60+ Models
| 1 |
[removed]
| 2025-04-17T22:17:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1pc51/1week_compare_chatgpt_deepseek_gemini_60_models/
|
Buffalo_Emotional
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1pc51
| false | null |
t3_1k1pc51
|
/r/LocalLLaMA/comments/1k1pc51/1week_compare_chatgpt_deepseek_gemini_60_models/
| false | false |
self
| 1 | null |
lmarena is a company now. which benchmark to trust?
| 1 |
[removed]
| 2025-04-17T22:21:53 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1pfsa/lmarena_is_a_company_now_which_benchmark_to_trust/
|
Unlikely-Okra3892
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1pfsa
| false | null |
t3_1k1pfsa
|
/r/LocalLLaMA/comments/1k1pfsa/lmarena_is_a_company_now_which_benchmark_to_trust/
| false | false |
self
| 1 | null |
Judging Embeddings
| 1 |
[removed]
| 2025-04-17T22:23:39 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1ph6z/judging_embeddings/
|
remyxai
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1ph6z
| false | null |
t3_1k1ph6z
|
/r/LocalLLaMA/comments/1k1ph6z/judging_embeddings/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'VQKIVyFaXVqS_-qOiQM9Mb5NKxNJ2ZmI_eeHhx4p3JA', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=108&crop=smart&auto=webp&s=55d30c9b3fb071b39044614c03cc1ecba14a48c7', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=216&crop=smart&auto=webp&s=fe482a54df69d9741c50bbcb894321457de762e0', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=320&crop=smart&auto=webp&s=837752a5c85c4963f7da3c75bf760afd14ac97b8', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=640&crop=smart&auto=webp&s=d7e456d129a7e0315ba3b18a9709904578fc92c4', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=960&crop=smart&auto=webp&s=310f85edae9aa2bbc19c1374c06bb4751461876e', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?width=1080&crop=smart&auto=webp&s=7cd9d1a856d474743894c143d5e6d2152af760fa', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/NsfImz-Hd-s9PVdkXwfQNA2FkfxC0lz1qdqM5PvFS70.jpg?auto=webp&s=9e4243400d6ee75b7bef46e9ca91421f9b9c7393', 'width': 1200}, 'variants': {}}]}
|
identify of riverhollow/riveroaks from lmarena
| 1 |
Looks like **riverhollow** is gone and I get similar output from **gemini-2.5-flash-preview-04-17**
Do you have some observations?
| 2025-04-17T22:24:30 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1phty/identify_of_riverhollowriveroaks_from_lmarena/
|
jacek2023
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1phty
| false | null |
t3_1k1phty
|
/r/LocalLLaMA/comments/1k1phty/identify_of_riverhollowriveroaks_from_lmarena/
| false | false |
self
| 1 | null |
Riverhollow from LMarena – Identity revealed?
| 0 |
Looks like **Riverhollow** from lmarena is gone and now I get similar output from **gemini-2.5-flash-preview-04-17**
What are your experiences? I was searching for the true name of it for days
| 2025-04-17T22:27:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1pjvz/riverhollow_from_lmarena_identity_revealed/
|
jacek2023
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1pjvz
| false | null |
t3_1k1pjvz
|
/r/LocalLLaMA/comments/1k1pjvz/riverhollow_from_lmarena_identity_revealed/
| false | false |
self
| 0 | null |
Gemini 2.5 Flash - First impressions
| 15 |
Google is rapidly evolving its Gemini models, and I recently got my hands on the preview versions designated as **Gemini 2.5 Flash** and **Gemini 2.5 Pro**.
Flash is positioned as the faster, more cost-effective option, while Pro targets peak performance, especially for complex reasoning. I put them head-to-head, particularly focusing on demanding tasks, and the results challenged the on-paper value proposition.
**The Pricing Picture (As Experienced):**
The per-token costs I encountered were:
* **Gemini 2.5 Flash (Preview):**
* Input: $0.15 / million tokens
* Output (Standard/"Non-Thinking"): $0.60 / million tokens
* Output ("Thinking Mode" - Implied High Usage Rate): $3.50 / million tokens
* **Gemini 2.5 Pro (Preview):**
* Input: $1.25 / million tokens
* Output: $10.00 / million tokens
**Performance & Thinking Quality: Flash's Achilles' Heel**
This is where the cost-effectiveness argument started to unravel for me. My focus was on the models' reasoning and problem-solving abilities.
* **Gemini 2.5 Flash's Thinking:** The quality of reasoning felt **very poor**. For complex problems requiring logical steps, its approach seemed inefficient and indirect. It struggled compared to the Pro version.
* **Token Inefficiency:** The most critical issue was Flash's token consumption. It consistently required **5-6 times more tokens** than **Gemini 2.5 Pro** to tackle the same task. The thinking process felt like it was deliberately burning tokens rather than finding the most direct solution path.
* **Subjective Benchmark:** I'd rate its reasoning quality slightly below a strong open-source model like Qwen-QWQ-32b.
**The Real-World Test: STEM Exam Problems**
To test this under pressure, I used tough STEM exam papers on both models.
* **Gemini 2.5 Pro (Preview):** Handled the problems with relative token efficiency for its reasoning process.
* **Gemini 2.5 Flash (Preview):** Despite its much lower per-token costs (even the $3.50 "thinking" rate vs Pro's $10.00), Flash **used vastly more tokens** for the same problems.
**The Bottom Line: Effective Cost vs. Sticker Price**
My conclusion based on these tests was clear: **For complex reasoning tasks, the preview version of Gemini 2.5 Flash effectively cost more per solved problem than the preview version of Gemini 2.5 Pro, despite Flash's lower per-token price.**
The extreme token inefficiency completely negated the cheaper rate. Paying $3.50 per million for Flash's "thinking" output tokens felt especially wasteful given the low quality and high volume required.
| 2025-04-17T22:43:03 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1pw3z/gemini_25_flash_first_impressions/
|
Embarrassed-Way-1350
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1pw3z
| false | null |
t3_1k1pw3z
|
/r/LocalLLaMA/comments/1k1pw3z/gemini_25_flash_first_impressions/
| false | false |
self
| 15 | null |
Generalized script for wakeword detection to run any script.
| 6 |
Wakeword: Generalized script that listens for a wakeword and runs a command you give it (so write a wrapper for your project that needs to be triggered with a wakeword):
#!/usr/bin/env python3
# by jaggz.h {who is at} gmail.com (and jaggzh on github)
# cc0
import asyncio
import time
import wave
import pvporcupine
import pyaudio
import struct
import io
import argparse
import subprocess
# models_basedir="~/wakegen/venv/lib/python3.11/site-packages/pvporcupine/resources/keyword_files/linux"
# alexa_linux.ppn grasshopper_linux.ppn picovoice_linux.ppn
# americano_linux.ppn 'hey google_linux.ppn' porcupine_linux.ppn
# blueberry_linux.ppn 'hey siri_linux.ppn' 'smart mirror_linux.ppn'
# bumblebee_linux.ppn jarvis_linux.ppn snowboy_linux.ppn
# computer_linux.ppn 'ok google_linux.ppn' terminator_linux.ppn
# grapefruit_linux.ppn 'pico clock_linux.ppn' 'view glass_linux.ppn'
# Configuration
DEF_KEYWORD_PATH = "~/wakegen/venv/lib/python3.11/site-packages/pvporcupine/resources/keyword_files/linux/blueberry_linux.ppn"
DEF_SENSITIVITY = 0.5 # Adjust sensitivity as needed
DEF_SR = 16000 # Sample rate of the audio
DEF_SAMPLE_WIDTH = 2 # Sample width of the audio
DEF_CHANNELS = 1 # Number of audio channels
DEF_RECORD_DURATION = .3 # Seconds to record
DEF_FRAME_LENGTH = 512 # Porcupine's frame length
# Initialize PyAudio
audio = pyaudio.PyAudio()
# Create Porcupine instance
porcupine = pvporcupine.create(
keyword_paths=[DEF_KEYWORD_PATH], sensitivities=[DEF_SENSITIVITY]
)
# Define function to record audio
async def record_audio(stream: pyaudio.Stream, frames_per_buffer: int):
"""Records audio for the specified duration."""
frames = []
start_time = time.time()
while time.time() - start_time < RECORD_DURATION:
data = stream.read(frames_per_buffer)
frames.append(data)
return b"".join(frames)
# Define function to process audio with Porcupine
async def process_audio(audio_data: bytes, cmd: str, non_blocking: bool):
"""Processes recorded audio with Porcupine and reports results."""
print("Processing audio... ", end='\r')
# Add WAV header
audio_data_with_header = add_wav_header(
audio_data, SAMPLE_RATE, SAMPLE_WIDTH, CHANNELS
)
# Now write the audio data with header
with wave.open(io.BytesIO(audio_data_with_header), "rb") as wf:
# Read audio in frames
for i in range(0, len(audio_data), FRAME_LENGTH * SAMPLE_WIDTH * CHANNELS):
frame_data = audio_data[i : i + FRAME_LENGTH * SAMPLE_WIDTH * CHANNELS]
# Unpack audio data into a list of samples
audio_samples = struct.unpack_from(
"h" * FRAME_LENGTH, frame_data
)
# Run Porcupine on the frame
keyword_index = porcupine.process(audio_samples)
if keyword_index >= 0:
print(f"Wake word detected! (Index: {keyword_index})")
if cmd:
print(f"Executing command: {cmd}")
try:
if non_blocking:
# Run command in the background
subprocess.Popen(cmd.split())
else:
# Run command and wait for it to finish
subprocess.run(cmd.split(), check=True)
except subprocess.CalledProcessError as e:
# Handle error if command execution fails
print(f"Command failed with error: {e}. Will try again next time.")
except Exception as e:
# Handle any other errors that might occur
print(f"An unexpected error occurred: {e}. Will try again next time.")
return # Exit after detection
print("Wake word not detected. ", end='\r')
async def main(keyword_path: str, sensitivity: float, sample_rate: int, sample_width: int, channels: int, record_duration: float, cmd: str, non_blocking: bool):
"""Main program loop."""
print("Listening for wake word...", end='\r')
global SAMPLE_RATE, SAMPLE_WIDTH, CHANNELS, RECORD_DURATION, FRAME_LENGTH
SAMPLE_RATE = sample_rate
SAMPLE_WIDTH = sample_width
CHANNELS = channels
RECORD_DURATION = record_duration
FRAME_LENGTH = porcupine.frame_length
# Create PyAudio stream
stream = audio.open(
format=pyaudio.paInt16,
channels=CHANNELS,
rate=SAMPLE_RATE,
input=True,
frames_per_buffer=FRAME_LENGTH,
)
while True:
# Record audio
audio_data = await record_audio(stream, FRAME_LENGTH)
# Process audio with Porcupine
await process_audio(audio_data, cmd, non_blocking)
# Close stream
stream.stop_stream()
stream.close()
def add_wav_header(audio_data: bytes, sample_rate: int, sample_width: int, channels: int):
"""Adds a WAV header to raw audio data."""
num_channels = channels
frame_rate = sample_rate
sample_width = sample_width
num_frames = len(audio_data) // (sample_width * num_channels)
# Compute audio data size
data_size = num_frames * num_channels * sample_width
# Create WAV header
header = b"RIFF"
header += struct.pack("<L", 36 + data_size) # Total file size
header += b"WAVE"
header += b"fmt "
header += struct.pack("<L", 16) # Length of fmt chunk
header += struct.pack("<H", 1) # Format code (1 for PCM)
header += struct.pack("<H", num_channels)
header += struct.pack("<L", frame_rate)
header += struct.pack("<L", frame_rate * num_channels * sample_width) # Byte rate
header += struct.pack("<H", num_channels * sample_width) # Block align
header += struct.pack("<H", sample_width * 8) # Bits per sample
header += b"data"
header += struct.pack("<L", data_size) # Size of data chunk
return header + audio_data
if __name__ == "__main__":
parser = argparse.ArgumentParser(prog="rhasspy-wake-porcupine-hermes")
parser.add_argument(
"-k",
"--keyword",
default=DEF_KEYWORD_PATH,
help="Path to Porcupine keyword file (.ppn)",
)
parser.add_argument(
"-s",
"--sensitivity",
type=float,
default=DEF_SENSITIVITY,
help="Sensitivity of keyword (default: 0.5)",
)
parser.add_argument(
"-r",
"--sample-rate",
type=int,
default=DEF_SR,
help=f"Sample rate of the audio (default: {DEF_SR})",
)
parser.add_argument(
"-w",
"--sample-width",
type=int,
default=DEF_SAMPLE_WIDTH,
help="Sample width of the audio (default: 2)",
)
parser.add_argument(
"-C",
"--channels",
type=int,
default=DEF_CHANNELS,
help="Number of audio channels (default: 1)",
)
parser.add_argument(
"-d",
"--record-duration",
type=float,
default=DEF_RECORD_DURATION,
help=f"Seconds to record audio (default: {DEF_RECORD_DURATION})",
)
parser.add_argument(
"-c",
"--cmd",
help="Command to execute when wake word is detected",
)
parser.add_argument(
"-B",
"--non-blocking",
action="store_true",
help="Run command in the background",
)
args = parser.parse_args()
# Recreate Porcupine with the provided keyword path and sensitivity
porcupine = pvporcupine.create(
keyword_paths=[args.keyword], sensitivities=[args.sensitivity]
)
asyncio.run(main(args.keyword, args.sensitivity, args.sample_rate, args.sample_width, args.channels, args.record_duration, args.cmd, args.non_blocking))
# Terminate PyAudio
audio.terminate()
| 2025-04-17T23:12:13 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1qiat/generalized_script_for_wakeword_detection_to_run/
|
jaggzh
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1qiat
| false | null |
t3_1k1qiat
|
/r/LocalLLaMA/comments/1k1qiat/generalized_script_for_wakeword_detection_to_run/
| false | false |
self
| 6 | null |
microsoft/MAI-DS-R1, DeepSeek R1 Post-Trained by Microsoft
| 332 | 2025-04-17T23:22:11 |
https://huggingface.co/microsoft/MAI-DS-R1
|
TKGaming_11
|
huggingface.co
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1qpr6
| false | null |
t3_1k1qpr6
|
/r/LocalLLaMA/comments/1k1qpr6/microsoftmaidsr1_deepseek_r1_posttrained_by/
| false | false | 332 |
{'enabled': False, 'images': [{'id': 'gNeLuucO5JeU3USiy3XqA0PJjXOcO7-aG3kRw83Z07c', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/oacNTVfe15Ozahhiv8YMZ-Teu__pPBVygtAgzE9FP3c.jpg?width=108&crop=smart&auto=webp&s=84981e30fc6e2f7d00268369cf58c3c1bca65504', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/oacNTVfe15Ozahhiv8YMZ-Teu__pPBVygtAgzE9FP3c.jpg?width=216&crop=smart&auto=webp&s=d2e19c29caf170d09b730606ff5e17c34b552fa2', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/oacNTVfe15Ozahhiv8YMZ-Teu__pPBVygtAgzE9FP3c.jpg?width=320&crop=smart&auto=webp&s=b9e149e7b6ccfa1b464248f0ecd48ad67ff34f4e', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/oacNTVfe15Ozahhiv8YMZ-Teu__pPBVygtAgzE9FP3c.jpg?width=640&crop=smart&auto=webp&s=140a805e8a294974bbef97d4e1035ef969130c5a', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/oacNTVfe15Ozahhiv8YMZ-Teu__pPBVygtAgzE9FP3c.jpg?width=960&crop=smart&auto=webp&s=cf8e496c6675ae3b2754218e6e3bae38ccfa29bf', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/oacNTVfe15Ozahhiv8YMZ-Teu__pPBVygtAgzE9FP3c.jpg?width=1080&crop=smart&auto=webp&s=f0e6df14eb3642ee00a80dcdc9d73cf1be0a481f', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/oacNTVfe15Ozahhiv8YMZ-Teu__pPBVygtAgzE9FP3c.jpg?auto=webp&s=cd3dac9f0a6c750e511ae05309175021aa37f32f', 'width': 1200}, 'variants': {}}]}
|
||
LLM distribution over Different OS
| 1 |
[removed]
| 2025-04-17T23:27:15 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1qthz/llm_distribution_over_different_os/
|
No_Draft_8756
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1qthz
| false | null |
t3_1k1qthz
|
/r/LocalLLaMA/comments/1k1qthz/llm_distribution_over_different_os/
| false | false |
self
| 1 | null |
I made this extension that applies the AI's changes semi-automatically without using an API.
| 14 |
Basically, the AI responds in a certain format, and when you paste it into the extension, it automatically executes the commands — creates files, etc. I made it in a short amount of time and wanted to know what you think. The idea was to have something that doesn't rely on APIs, which usually have a lot of limitations. It can be used with any AI — you just need to set the system instructions.
If I were to continue developing it, I'd add more efficient editing (without needing to show the entire code), using search and replace, and so on.
[https://marketplace.visualstudio.com/items/?itemName=FelpolinColorado.buildy](https://marketplace.visualstudio.com/items/?itemName=FelpolinColorado.buildy)
**LIMITATIONS AND WARNING:** this extension is not secure at all. Even though it has a checkpoint system, it doesn’t ask for any permissions, so be very careful if you choose to use it.
| 2025-04-17T23:57:18 |
https://v.redd.it/jxyur24afhve1
|
Delicious-Trash6988
|
/r/LocalLLaMA/comments/1k1rfe7/i_made_this_extension_that_applies_the_ais/
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1rfe7
| false |
{'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/jxyur24afhve1/DASHPlaylist.mpd?a=1747655845%2CMGViYTEyNTY0ZjNmNjM3NjBlZTAyODZjODE4MDJjZDMzODExODM1ZmRjOTMxNDkwNDM5OTdjYTQ5MTQ2ZmYzYw%3D%3D&v=1&f=sd', 'duration': 86, 'fallback_url': 'https://v.redd.it/jxyur24afhve1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/jxyur24afhve1/HLSPlaylist.m3u8?a=1747655845%2CZGM3MTgzYjhiYzNhM2Y2OWU1OWUzNTI3ODU1MTY2ZmMwNTUxMzE2NjU4YjY1YjhlMTYyZTMwODM5NGNjMzA3OQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/jxyur24afhve1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}}
|
t3_1k1rfe7
|
/r/LocalLLaMA/comments/1k1rfe7/i_made_this_extension_that_applies_the_ais/
| false | false | 14 |
{'enabled': False, 'images': [{'id': 'bnM5aHUyNGFmaHZlMfXWszA2aBwWyFSKhA5ZAJVyEflwSFKnBokwQQKV24Gs', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/bnM5aHUyNGFmaHZlMfXWszA2aBwWyFSKhA5ZAJVyEflwSFKnBokwQQKV24Gs.png?width=108&crop=smart&format=pjpg&auto=webp&s=9d720fba9105a9a7e8a448689d04d630a452aead', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/bnM5aHUyNGFmaHZlMfXWszA2aBwWyFSKhA5ZAJVyEflwSFKnBokwQQKV24Gs.png?width=216&crop=smart&format=pjpg&auto=webp&s=7e4441cd18583db531cea105beb9f46744dbfcf3', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/bnM5aHUyNGFmaHZlMfXWszA2aBwWyFSKhA5ZAJVyEflwSFKnBokwQQKV24Gs.png?width=320&crop=smart&format=pjpg&auto=webp&s=6d92c1dd6a4e664004696dfa7a6abf5bf9346501', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/bnM5aHUyNGFmaHZlMfXWszA2aBwWyFSKhA5ZAJVyEflwSFKnBokwQQKV24Gs.png?width=640&crop=smart&format=pjpg&auto=webp&s=76fc0d0131ae84c36e9e59db7118622e748d52c8', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/bnM5aHUyNGFmaHZlMfXWszA2aBwWyFSKhA5ZAJVyEflwSFKnBokwQQKV24Gs.png?width=960&crop=smart&format=pjpg&auto=webp&s=901644c6723af566aaac70667a8a708de9027fc8', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/bnM5aHUyNGFmaHZlMfXWszA2aBwWyFSKhA5ZAJVyEflwSFKnBokwQQKV24Gs.png?width=1080&crop=smart&format=pjpg&auto=webp&s=a1b9a36be3b5eda88494bbc63f7314cb104f6b25', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/bnM5aHUyNGFmaHZlMfXWszA2aBwWyFSKhA5ZAJVyEflwSFKnBokwQQKV24Gs.png?format=pjpg&auto=webp&s=a7a0708b57d3ad486332091a7e2173370e3a72ed', 'width': 1920}, 'variants': {}}]}
|
|
How to run Llama 4 fast, even though it's too big to fit in RAM
| 108 |
TL;DR: in your llama.cpp command, add:
`-ngl 49 --override-tensor "([0-9]+).ffn_.*_exps.=CPU" --ubatch-size 1`
Explanation:
`-ngl 49`
* offload all 49 layers to GPU
`--override-tensor "([0-9]+).ffn_.*_exps.=CPU"`
* ...except for the MOE weights
`--ubatch-size 1`
* process the prompt in batches of 1 at a time (instead of the default 512 - otherwise your SSD will be the bottleneck and prompt processing will be slower)
This radically speeds up inference by taking advantage of LLama 4's MOE architecture. LLama 4 Maverick has 400 billion total parameters, but only 17 billion active parameters. Some are needed on every token generation, while others are only occasionally used. So if we put the parameters that are always needed onto GPU, those will be processed quickly, and there will just be a small number that need to be handled by the CPU. This works so well that the weights don't even need to all fit in your CPU's RAM - many of them can memory mapped from NVMe.
My results with Llama 4 Maverick:
* Unsloth's UD-Q4\_K\_XL quant is 227GB
* Unsloth's Q8\_0 quant is 397GB
Both of those are much bigger than my RAM + VRAM (128GB + 3x24GB). But with these tricks, I get 15 tokens per second with the UD-Q4\_K\_M and 6 tokens per second with the Q8\_0.
Full llama.cpp server commands:
Note: the `--override-tensor` command is tweaked because I had some extra VRAM available, so I offloaded most of the MOE layers to CPU, but loaded a few onto each GPU.
UD-Q4\_K\_XL:
`./llama-server -m Llama-4-Maverick-17B-128E-Instruct-UD-Q4_K_XL-00001-of-00005.gguf -ngl 49 -fa -c 16384 --override-tensor "([1][1-9]|[2-9][0-9]).ffn_.*_exps.=CPU,([0-2]).ffn_.*_exps.=CUDA0,([3-6]).ffn_.*_exps.=CUDA1,([7-9]|[1][0]).ffn_.*_exps.=CUDA2" --ubatch-size 1`
Q8\_0:
`./llama-server -m Llama-4-Maverick-17B-128E-Instruct-Q8_0-00001-of-00009.gguf -ngl 49 -fa -c 16384 --override-tensor "([6-9]|[1-9][0-9]).ffn_.*_exps.=CPU,([0-1]).ffn_.*_exps.=CUDA0,([2-3]).ffn_.*_exps.=CUDA1,([4-5]).ffn_.*_exps.=CUDA2" --ubatch-size 1`
Credit goes to the people behind Unsloth for this knowledge. I hadn't seen people talking about this here, so I thought I'd make a post.
| 2025-04-18T00:03:01 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1rjm1/how_to_run_llama_4_fast_even_though_its_too_big/
|
Klutzy-Snow8016
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1rjm1
| false | null |
t3_1k1rjm1
|
/r/LocalLLaMA/comments/1k1rjm1/how_to_run_llama_4_fast_even_though_its_too_big/
| false | false |
self
| 108 | null |
What are your favorite Tiny models?
| 1 |
[removed]
| 2025-04-18T00:38:47 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1s8nw/what_are_your_favorite_tiny_models/
|
Radhe_sito
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1s8nw
| false | null |
t3_1k1s8nw
|
/r/LocalLLaMA/comments/1k1s8nw/what_are_your_favorite_tiny_models/
| false | false |
self
| 1 | null |
What are your favorite Tiny models?
| 1 |
[removed]
| 2025-04-18T00:39:53 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1s9g0/what_are_your_favorite_tiny_models/
|
Radhe_sito
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1s9g0
| false | null |
t3_1k1s9g0
|
/r/LocalLLaMA/comments/1k1s9g0/what_are_your_favorite_tiny_models/
| false | false |
self
| 1 | null |
Local LLM beginner here - a question about best models to use for my scenario
| 2 |
So I've only briefly dabbled into running LLMs locally, I have Ollama setup, and run a couple versions of the deepseek-r1 model.
That's all my background for local LLMs. So I'm curious what would be best for my scenario.
I downloaded all of my account's reddit data, past comments and posts. I want to create some kind of local model that uses the comments as training data, and enact my reddit persona.
What local models or processes would work best for this?
| 2025-04-18T00:40:44 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1sa1t/local_llm_beginner_here_a_question_about_best/
|
VaderOnReddit
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1sa1t
| false | null |
t3_1k1sa1t
|
/r/LocalLLaMA/comments/1k1sa1t/local_llm_beginner_here_a_question_about_best/
| false | false |
self
| 2 | null |
What are your favorite Tiny models?
| 1 |
[removed]
| 2025-04-18T00:45:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1sd5x/what_are_your_favorite_tiny_models/
|
Radhe_sito
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1sd5x
| false | null |
t3_1k1sd5x
|
/r/LocalLLaMA/comments/1k1sd5x/what_are_your_favorite_tiny_models/
| false | false |
self
| 1 | null |
Which is best LLM for academic writing now? (esp. for non-native English speakers)
| 1 |
[removed]
| 2025-04-18T01:00:39 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1sny5/which_is_best_llm_for_academic_writing_now_esp/
|
SpinachAn
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1sny5
| false | null |
t3_1k1sny5
|
/r/LocalLLaMA/comments/1k1sny5/which_is_best_llm_for_academic_writing_now_esp/
| false | false |
self
| 1 | null |
LMSA - LM Studio Assistant for Android
| 1 |
[removed]
| 2025-04-18T01:13:02 |
https://lmsa.app
|
CowMan30
|
lmsa.app
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1swjp
| false | null |
t3_1k1swjp
|
/r/LocalLLaMA/comments/1k1swjp/lmsa_lm_studio_assistant_for_android/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'Ks-o3a_WQH9p4rGA6yWXbTAoSyeg-D_HLOJhQqW8ZCA', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/Y_vCaQDJl5lK4_piimMu9SrLo1vTfhOy7FqxgSPjKDA.jpg?width=108&crop=smart&auto=webp&s=36bd57b9babff558165d6d9095acf73a00f23575', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/Y_vCaQDJl5lK4_piimMu9SrLo1vTfhOy7FqxgSPjKDA.jpg?width=216&crop=smart&auto=webp&s=82a25cb976cf0ddfef02561520cfd2c4ea6402dc', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/Y_vCaQDJl5lK4_piimMu9SrLo1vTfhOy7FqxgSPjKDA.jpg?width=320&crop=smart&auto=webp&s=d83e21f72447e9e1a7e1115751e9f86b496739a9', 'width': 320}], 'source': {'height': 512, 'url': 'https://external-preview.redd.it/Y_vCaQDJl5lK4_piimMu9SrLo1vTfhOy7FqxgSPjKDA.jpg?auto=webp&s=c1c58057b58cfa032b1461824f937138b821bb3e', 'width': 512}, 'variants': {}}]}
|
|
Mac VRAM Allocation taskbar app VRAM Pro
| 2 |
I built a tiny macOS utility that does one very specific thing:
**It unlocks additional GPU memory on Apple Silicon Macs.**
Why? Because macOS doesn’t give you any control over VRAM — even though you *technically* have the bandwidth.
I needed it for smoother performance in:
* Running large LLMs
* Blender and After Effects
* Unity and Unreal previews
So… I made **VRAM Pro**.
It’s:
* 🧠 Simple: Just sits in your menubar
* 🔓 Lets you allocate more VRAM (temporarily and safely)
* 🔐 Notarized, signed, Sparkle-updated
* 💸 Totally free (I just built it for fun and needed it)
# 📦 Download:
[https://VRAMPro.com](https://VRAMPro.com)
Would love feedback, and happy to tweak it based on use cases!
Also — if you’ve got other obscure GPU tricks on macOS, I’d *love* to hear them.
Thanks Reddit 🙏
PS: after I made this app someone created am open source copy: [https://github.com/PaulShiLi/Siliv](https://github.com/PaulShiLi/Siliv)
| 2025-04-18T01:24:53 |
https://www.reddit.com/gallery/1k1t4pa
|
DazzlingHedgehog6650
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1t4pa
| false | null |
t3_1k1t4pa
|
/r/LocalLLaMA/comments/1k1t4pa/mac_vram_allocation_taskbar_app_vram_pro/
| false | false | 2 | null |
|
Instantly allocate more graphics memory on your Mac VRAM Pro
| 38 |
I built a tiny macOS utility that does one very specific thing:
**It unlocks additional GPU memory on Apple Silicon Macs.**
Why? Because macOS doesn’t give you any control over VRAM — and hard caps it, leading to swap issues in certain use cases.
I needed it for performance in:
* Running large LLMs
* Blender and After Effects
* Unity and Unreal previews
So… I made **VRAM Pro**.
It’s:
* 🧠 Simple: Just sits in your menubar
* 🔓 Lets you allocate more VRAM
* 🔐 Notarized, signed, autoupdates
# 📦 Download:
[https://VRAMPro.com](https://vrampro.com/)
Do you need this app? No! You can do this with various commands in terminal. But wanted a nice and easy GUI way to do this.
Would love feedback, and happy to tweak it based on use cases!
Also — if you’ve got other obscure GPU tricks on macOS, I’d *love* to hear them.
Thanks Reddit 🙏
PS: after I made this app someone created am open source copy: [https://github.com/PaulShiLi/Siliv](https://github.com/PaulShiLi/Siliv)
| 2025-04-18T01:38:15 |
https://www.reddit.com/gallery/1k1tdpa
|
DazzlingHedgehog6650
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1tdpa
| false | null |
t3_1k1tdpa
|
/r/LocalLLaMA/comments/1k1tdpa/instantly_allocate_more_graphics_memory_on_your/
| false | false | 38 | null |
|
How to pick when to use which?
| 0 |
With so many models releasing per week. Is there any single interface other then Ollama for also paid models that you guys use to decide when to use which showing benchmarking, type of data etc etc?
| 2025-04-18T01:51:09 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1tmay/how_to_pick_when_to_use_which/
|
semicolon-10
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1tmay
| false | null |
t3_1k1tmay
|
/r/LocalLLaMA/comments/1k1tmay/how_to_pick_when_to_use_which/
| false | false |
self
| 0 | null |
Ollama version history, change listings, backwards-compatibility
| 0 |
Apologies for the probably dumb question, but where on GitHub does one look to get a detailed description of what has changed from one version of Ollama to the next? I ask because I went to load a new model and received a message of "*The model you are attempting to pull requires a newer version of Ollam*a." I am hesitant to change anything that might break the model setups I have running now which are running perfectly, but I also want to feel more confident moving forward in adding new models, and to see if Ollama has added any functionality that I may have manually coded for that I may not need to going forward... Thanks for your direction & guideance.
| 2025-04-18T01:56:43 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1tq2b/ollama_version_history_change_listings/
|
DelosBoard2052
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1tq2b
| false | null |
t3_1k1tq2b
|
/r/LocalLLaMA/comments/1k1tq2b/ollama_version_history_change_listings/
| false | false |
self
| 0 | null |
I cannot seem to be able to use the YuE Music Ai in Pinokio
| 1 |
[removed]
| 2025-04-18T01:59:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1trl7/i_cannot_seem_to_be_able_to_use_the_yue_music_ai/
|
Select_Rutabaga7384
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1trl7
| false | null |
t3_1k1trl7
|
/r/LocalLLaMA/comments/1k1trl7/i_cannot_seem_to_be_able_to_use_the_yue_music_ai/
| false | false |
self
| 1 | null |
New to all of this. Any point to adding a second card with 1GB GDDR5? Have RTX 4080, up for planning to buy more DDR5.
| 1 |
[removed]
| 2025-04-18T02:21:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1u6r7/new_to_all_of_this_any_point_to_adding_a_second/
|
17usc
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1u6r7
| false | null |
t3_1k1u6r7
|
/r/LocalLLaMA/comments/1k1u6r7/new_to_all_of_this_any_point_to_adding_a_second/
| false | false |
self
| 1 | null |
Attempting to Solve the Cross-Platform AI Billing Challenge as a Solo Engineer/Founder - Need Feedback
| 1 |
[removed]
| 2025-04-18T02:30:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1uchw/attempting_to_solve_the_crossplatform_ai_billing/
|
Frozen-Insightful-22
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1uchw
| false | null |
t3_1k1uchw
|
/r/LocalLLaMA/comments/1k1uchw/attempting_to_solve_the_crossplatform_ai_billing/
| false | false |
self
| 1 | null |
It's been hours already without a new open source SOTA being released. Is open source LLMs dead? Is this it???
| 337 |
RIP LLMS!
| 2025-04-18T02:40:50 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1ujk4/its_been_hours_already_without_a_new_open_source/
|
DamiaHeavyIndustries
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1ujk4
| false | null |
t3_1k1ujk4
|
/r/LocalLLaMA/comments/1k1ujk4/its_been_hours_already_without_a_new_open_source/
| false | false |
self
| 337 | null |
Multilingual pretraining datasets
| 3 |
I’m planning to continuous retrain multilingual models and would love to know which multilingual pretraining datasets are available on Hugging Face. Can anyone share some suggestions or links to datasets that cover multiple languages?
Thanks in advance!
| 2025-04-18T02:56:20 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1utq4/multilingual_pretraining_datasets/
|
MarySmith2021
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1utq4
| false | null |
t3_1k1utq4
|
/r/LocalLLaMA/comments/1k1utq4/multilingual_pretraining_datasets/
| false | false |
self
| 3 | null |
Embedding Model Recommended for RAG
| 1 |
[removed]
| 2025-04-18T03:16:43 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1v6pt/embedding_model_recommended_for_rag/
|
rich188
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1v6pt
| false | null |
t3_1k1v6pt
|
/r/LocalLLaMA/comments/1k1v6pt/embedding_model_recommended_for_rag/
| false | false |
self
| 1 | null |
RAG/Embedding Model Suggestion
| 1 |
[removed]
| 2025-04-18T03:17:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1v78j/ragembedding_model_suggestion/
|
rich188
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1v78j
| false | null |
t3_1k1v78j
|
/r/LocalLLaMA/comments/1k1v78j/ragembedding_model_suggestion/
| false | false |
self
| 1 | null |
CSM 1B is real-time now and has fine-tuning
| 187 |
[https://github.com/davidbrowne17/csm-streaming](https://github.com/davidbrowne17/csm-streaming)
Not sure if many of you have been following this model, but the open-source community has managed to reach real-time with streaming and figured out fine-tuning. This is my repo with fine-tuning and a chat demo, my version of fine-tuning is lora but there is also full fine tuning out there as well. Give it a try and let me know how it compares to other TTS models.
| 2025-04-18T03:21:15 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1v9rq/csm_1b_is_realtime_now_and_has_finetuning/
|
SovietWarBear17
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1v9rq
| false | null |
t3_1k1v9rq
|
/r/LocalLLaMA/comments/1k1v9rq/csm_1b_is_realtime_now_and_has_finetuning/
| false | false |
self
| 187 |
{'enabled': False, 'images': [{'id': 'LCyVaSFQLCAx8kjuQVBBo2_vWEBDGqxRFTFz6tdmxj8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/gb2LuERauBVYmlwqvVwLQOzt-W6J1LSgxmnCHRm7KhI.jpg?width=108&crop=smart&auto=webp&s=fbad08d22911d2dcef13553751375e098f952a11', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/gb2LuERauBVYmlwqvVwLQOzt-W6J1LSgxmnCHRm7KhI.jpg?width=216&crop=smart&auto=webp&s=105283a2792f4c0a5c006516982fcf305cf43eed', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/gb2LuERauBVYmlwqvVwLQOzt-W6J1LSgxmnCHRm7KhI.jpg?width=320&crop=smart&auto=webp&s=fbb333f746520b10eb9b884ea292a4fa64404e2a', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/gb2LuERauBVYmlwqvVwLQOzt-W6J1LSgxmnCHRm7KhI.jpg?width=640&crop=smart&auto=webp&s=e182954dc37a82f21efb78c50e0587540c5e540e', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/gb2LuERauBVYmlwqvVwLQOzt-W6J1LSgxmnCHRm7KhI.jpg?width=960&crop=smart&auto=webp&s=450554a908fdc33047345d839733e757899ef4bd', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/gb2LuERauBVYmlwqvVwLQOzt-W6J1LSgxmnCHRm7KhI.jpg?width=1080&crop=smart&auto=webp&s=cc8a72e3ce7da72cabe25496acea498bfb597abe', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/gb2LuERauBVYmlwqvVwLQOzt-W6J1LSgxmnCHRm7KhI.jpg?auto=webp&s=4550789df7ae5f9b892fc8f1af8979d634890ae8', 'width': 1200}, 'variants': {}}]}
|
Best 7b-14b models for roleplaying?
| 7 |
What are some of the best uncensored models to run with 12gb of VRAM that work good for roleplaying?
| 2025-04-18T03:44:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1vobv/best_7b14b_models_for_roleplaying/
|
AsDaylight_Dies
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1vobv
| false | null |
t3_1k1vobv
|
/r/LocalLLaMA/comments/1k1vobv/best_7b14b_models_for_roleplaying/
| false | false |
self
| 7 | null |
No API keys, no cloud. Just local Al + tools that actually work. Too much to ask?
| 129 |
It's been about a month since we first posted Clara here.
Clara is a local-first Al assistant - think of it like ChatGPT, but fully private and running on your own machine using Ollama.
Since the initial release, I've had a small group of users try it out, and I've pushed several updates based on real usage and feedback.
The biggest update is that Clara now comes with n8n built-in.
That means you can now build and run your own tools directly inside the assistant - no setup needed, no external services. Just open Clara and start automating.
With the n8n integration, Clara can now do more than chat. You can use it to:
• Check your emails
• Manage your calendar
• Call APIs
• Run scheduled tasks
• Process webhooks
• Connect to databases
• And anything else you can wire up using n8n's visual flow builder
The assistant can trigger these workflows directly - so you can talk to Clara and ask it to do real tasks, using tools that run entirely on your
device.
Everything happens locally. No data goes out, no accounts, no cloud dependency.
If you're someone who wants full control of your Al and automation setup, this might be something worth trying.
You can check out the project here:
GitHub: https://github.com/badboysm890/ClaraVerse
Thanks to everyone who's been trying it and sending feedback. Still improving things - more updates soon.
Note: I'm aware of great projects like OpenWebUI and LibreChat. Clara takes a slightly different approach - focusing on reducing dependencies, offering a native desktop app, and making the overall experience more user-friendly so that more people can easily get started with local Al.
| 2025-04-18T03:56:37 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1vvy3/no_api_keys_no_cloud_just_local_al_tools_that/
|
aruntemme
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1vvy3
| false | null |
t3_1k1vvy3
|
/r/LocalLLaMA/comments/1k1vvy3/no_api_keys_no_cloud_just_local_al_tools_that/
| false | false |
self
| 129 |
{'enabled': False, 'images': [{'id': 'p7ljLQU7bp3E2GIpmDZ67UksN7JuGQCQ2apWw-NPzRQ', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/Bv5S_BWRPsbNpVHFb4PUPcV-pnozuaMHiCZ4G3Q8ICA.jpg?width=108&crop=smart&auto=webp&s=27d0b23a9a4d3110c37c9c8f8cfab6e328d6db86', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/Bv5S_BWRPsbNpVHFb4PUPcV-pnozuaMHiCZ4G3Q8ICA.jpg?width=216&crop=smart&auto=webp&s=152edc76f543e5729d3c491d2d4542b4b499b152', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/Bv5S_BWRPsbNpVHFb4PUPcV-pnozuaMHiCZ4G3Q8ICA.jpg?width=320&crop=smart&auto=webp&s=1ddd9ee0afc6041982c40c39b0c35686aea07128', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/Bv5S_BWRPsbNpVHFb4PUPcV-pnozuaMHiCZ4G3Q8ICA.jpg?width=640&crop=smart&auto=webp&s=0c582936f7ae5dcb6e9ed0a553a27bd35295a65c', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/Bv5S_BWRPsbNpVHFb4PUPcV-pnozuaMHiCZ4G3Q8ICA.jpg?width=960&crop=smart&auto=webp&s=8ab983dc6b66ad6ab70d3d8a84ae52e5bb920348', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/Bv5S_BWRPsbNpVHFb4PUPcV-pnozuaMHiCZ4G3Q8ICA.jpg?width=1080&crop=smart&auto=webp&s=03928a5fb24857a64df8bcd05443e8b10f3f344b', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/Bv5S_BWRPsbNpVHFb4PUPcV-pnozuaMHiCZ4G3Q8ICA.jpg?auto=webp&s=262e747d7295bc13965d3d5a35f04ebad6c7d3a0', 'width': 1200}, 'variants': {}}]}
|
If you had a time machine and went back 10 years in the past armed only with your laptop with some local ai on it. How could you use it to make money?
| 0 |
One easy technique would be to write articles as a freelancer.
| 2025-04-18T05:24:56 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1xc05/if_you_had_a_time_machine_and_went_back_10_years/
|
ImaginaryRea1ity
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1xc05
| false | null |
t3_1k1xc05
|
/r/LocalLLaMA/comments/1k1xc05/if_you_had_a_time_machine_and_went_back_10_years/
| false | false |
self
| 0 | null |
vLLM with transformers backend
| 56 |
You can try out the new integration with which you can run ANY transformers model with vLLM (even if it is not natively supported by vLLM)
Read more about it here: [https://blog.vllm.ai/2025/04/11/transformers-backend.html](https://blog.vllm.ai/2025/04/11/transformers-backend.html)
What can one do with this:
1. 1. Read the blog 😌
2. 2. Contribute to transformers - making models vLLM compatible
3. 3. Raise issues if you spot a bug with the integration
Vision Language Model support is coming very soon! Until any further announcements, we would love for everyone to stick using this integration with text only models 🤗
| 2025-04-18T05:32:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1xgco/vllm_with_transformers_backend/
|
Disastrous-Work-1632
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1xgco
| false | null |
t3_1k1xgco
|
/r/LocalLLaMA/comments/1k1xgco/vllm_with_transformers_backend/
| false | false |
self
| 56 |
{'enabled': False, 'images': [{'id': '-oF_CESVxWGYmnXEeT_RMEAFXNIW4qe8IPBbPGyqdrY', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/wgHpfpaqH4e5VLCswLxIoCAioXF-UEidtLP_BvomJ_U.jpg?width=108&crop=smart&auto=webp&s=9d0de14b8d3744134afab6fa341a4bbef35e5d28', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/wgHpfpaqH4e5VLCswLxIoCAioXF-UEidtLP_BvomJ_U.jpg?width=216&crop=smart&auto=webp&s=91e2fc49d9a6eb3974fbf79b92a937b4283f07d0', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/wgHpfpaqH4e5VLCswLxIoCAioXF-UEidtLP_BvomJ_U.jpg?width=320&crop=smart&auto=webp&s=ac463ad0def021662f135cfe8e31edbc2c0e9350', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/wgHpfpaqH4e5VLCswLxIoCAioXF-UEidtLP_BvomJ_U.jpg?width=640&crop=smart&auto=webp&s=f1c5f3722aee60cc6578ffa6925b2883590d7cec', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/wgHpfpaqH4e5VLCswLxIoCAioXF-UEidtLP_BvomJ_U.jpg?width=960&crop=smart&auto=webp&s=26fd2013f38f300801487544dba88de4cbab2c23', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/wgHpfpaqH4e5VLCswLxIoCAioXF-UEidtLP_BvomJ_U.jpg?width=1080&crop=smart&auto=webp&s=6053880162669a44009fd67cdb89349b45d60130', 'width': 1080}], 'source': {'height': 650, 'url': 'https://external-preview.redd.it/wgHpfpaqH4e5VLCswLxIoCAioXF-UEidtLP_BvomJ_U.jpg?auto=webp&s=513edc635fac5581cfbe8d8af6f8d7fdd50b1ece', 'width': 1300}, 'variants': {}}]}
|
Gemma3 VS Qwen2.5 math.
| 5 |
I manually did the math to calculate the highest voltage a solar panel in my array can be using temperature coefficient.
46.50 + (1-21-45) x (-.31 x 46.50/100)
the correct answer is 55.87
I wanted to make sure my math was right so I asked.
\### query
46.50 + (1-21-45) x (-.31 x 46.50/100)
use two decimal points.
\###
gemma3:4b-it-q4\_K\_M
Let's break down the calculation step by step:
1. **(1 - 21 - 45)**: 1 - 21 - 45 = -65
2. **(-.31 x 46.50 / 100)**: -.31 \* 46.50 = -14.415 -14.415 / 100 = -0.14415
3. **(1 - 21 - 45) x (-.31 x 46.50 / 100)**: -65 \* (-0.14415) = 9.25325
4. **46.50 + (1 - 21 - 45) x (-.31 x 46.50 / 100)**: 46.50 + 9.25325 = 55.75325
Rounded to two decimal places, the result is 55.75.
Therefore, the answer is **55.75**
\###
gemma3:4b-it-q8\_0
Okay, let's break down this calculation step by step:
1. **Calculate the parentheses first:** (1 - 21 - 45) = -65
2. **Calculate the multiplication within the parentheses:** \-0.31 \* 46.50 = 14.565
3. **Divide by 100:** 14.565 / 100 = 0.14565
4. **Multiply by -1:** \-0.14565 \* -1 = 0.14565
5. **Add the entire expression:** 46.50 + 0.14565 = 46.64565
6. **Round to two decimal places:** 46.64565 rounds to 46.65
Therefore, 46.50 + (1-21-45) x (-.31 x 46.50 /100) = **46.65**
\###
gemma3:12b-it-q4\_K\_M
Okay, let's break this down step by step following the order of operations (PEMDAS/BODMAS):
1. **Parentheses/Brackets first:** (1-21-45)
* 1 - 21 = -20
* \-20 - 45 = -65
1. **Inside the remaining parentheses:** (-.31 x 46.50)
* \-0.31 x 46.50 = -14.385
1. **Division:** \-14.385 / 100
* \-14.385 / 100 = -0.14385
1. **Multiplication:** \-65 x -0.14385
* \-65 \* -0.14385 = 9.34925
2. **Addition:** 46.50 + 9.34925
* 46.50 + 9.34925 = 55.84925
1. **Round to two decimal places:** 55.84925 rounds to **55.85**
Therefore, the answer is **55.85**
\###
gemma3:12b-it-q8\_0
Okay, let's break this down step-by-step, following the order of operations (PEMDAS/BODMAS):
1. **Parentheses/Brackets:**
* Inside the first set of parentheses: 1 - 21 - 45 = -65
* Inside the second set of parentheses: -.31 x 46.50 = -14.385
* Then, divide the result of the second set of parentheses by 100: -14.385 / 100 = -0.14385
2. **Multiplication:**
* (-65) x (-0.14385) = 9.35025
3. **Addition:**
* 46.50 + 9.35025 = 55.85025
4. **Round to two decimal places:**
* 55.85
**Therefore, the answer is 55.85**
\###
gemma3:27b-it-q4\_K\_M
Here's the calculation with two decimal points:
1. **Parentheses first:** (1 - 21 - 45) = -65
2. **Multiplication within parentheses:** \-0.31 \* 46.50 = -14.415
3. **Division:** \-14.415 / 100 = -0.14415
4. **Multiplication:** \-65 \* -0.14415 = 9.36975
5. **Addition:** 46.50 + 9.36975 = 55.86975
Rounding to two decimal places, the answer is **55.87**. <<<--- Bingo!
\###
qwen2.5:3b-instruct-q4\_K\_MToday at 1:09 AM
To solve the expression 46.50+(1−21−45)×(−.31×46.50/100), we need to follow the order of operations, which is Parentheses, Exponents, Multiplication and Division (from left to right), Addition and Subtraction (from left to right).
First, let's simplify inside the parentheses:
1. Calculate (1−21−45):1−21=−20−20−45=−65
So, (1−21−45)=−65.
Next, calculate the expression inside the second set of parentheses:
1. Calculate −.31×46.50/100:−.31×46.50=−14.415−14.415/100=−0.14415
Now, multiply this result by the value inside the first set of parentheses:
1. Multiply −65 by −0.14415:−65×−.14415=9.48975
Finally, add this to 46.50:
1. Add the results:46.50+9.48975=55.98975
Rounding to two decimal points, we get 55.99.
So, the final answer is **55.99**.
\###
qwen2.5:3b-instruct-q8\_0
To solve the expression 46.50+(1−21−45)×(−0.31×46.50/100), let's break it down step by step.
First, calculate inside the parentheses:
1−21−45=−65
Next, calculate the expression within the second set of parentheses:
−0.31×46.50/100
To do this:
−0.31×46.50=−14.415
Then divide by 100:
−14.415/100=−0.14415
Now, multiply the result from above by −65:
−65×(−0.14415)=9.36725
Finally, add this to 46.50:
46.50+9.36725=55.86725
Rounded to two decimal points, the result is:
55.87 <<<-- Bingo, wait a minute, what? Really? 3B?
\##################################################
it took Gemma3-27B to get this correct and Qwen2.5 did it with a 3B model? Really?
What in the heck is going on here?
I have always used qwen2.5 but when Gemma 3 arrived, I decided to take it for a spin and used it for while and for a lot of stuff it is pretty good but qwen2.5 destroyed Gemma 3 here. Has anyone seen similar?
I know LLM's are known to be bad a math but this kind of puts Qwen2.5 on top of my list as it seems to be great at pretty much everything I have thrown at it. Anything better than Qwen2.5 at a reasonable size like under 32B? Or is Qwen2.5 still king of the hill at 32B or below?
| 2025-04-18T05:39:10 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1xjsu/gemma3_vs_qwen25_math/
|
Boricua-vet
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1xjsu
| false | null |
t3_1k1xjsu
|
/r/LocalLLaMA/comments/1k1xjsu/gemma3_vs_qwen25_math/
| false | false |
self
| 5 | null |
I want to know if its possible to run a llama model in a old CPU.
| 2 |
I'm new to using Llama and I'd like to know if there are super lightweight models that can run on weak system's.
The system spec in question:
Intel(R) Pentium(R) Silver N6005 @ 2.00GHz, 1997 Mhz, 4 Core(s), 4 Logical Processor(s),with 16 GB ram.
| 2025-04-18T05:53:47 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1xrni/i_want_to_know_if_its_possible_to_run_a_llama/
|
Hoshino_Ruby
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1xrni
| false | null |
t3_1k1xrni
|
/r/LocalLLaMA/comments/1k1xrni/i_want_to_know_if_its_possible_to_run_a_llama/
| false | false |
self
| 2 | null |
Where is the promised open Grok 2?
| 210 |
As far as I know, Grok 2 was supposed to be open-sourced some time after Grok 3's release. But I'm afraid that by the time they decide to open-source Grok 2, it will already be completely obsolete. This is because even now, it significantly lags behind in performance compared to the likes of DeepSeek V3, and we also have Qwen 3 and Llama 4 Reasoning on the horizon (not to mention a potential open model from OpenAI). I believe that when they eventually decide to release it to the community, it will be of no use to anyone anymore, much like what happened with Grok 1. What are your thoughts on this?
| 2025-04-18T06:01:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1xvvr/where_is_the_promised_open_grok_2/
|
AlexBefest
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1xvvr
| false | null |
t3_1k1xvvr
|
/r/LocalLLaMA/comments/1k1xvvr/where_is_the_promised_open_grok_2/
| false | false |
self
| 210 | null |
Stop chasing agent frameworks. A mental model on separating out the low-level from high-level logic.
| 1 |
[removed]
| 2025-04-18T06:20:49 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1y68l/stop_chasing_agent_frameworks_a_mental_model_on/
|
AdditionalWeb107
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1y68l
| false | null |
t3_1k1y68l
|
/r/LocalLLaMA/comments/1k1y68l/stop_chasing_agent_frameworks_a_mental_model_on/
| false | false |
self
| 1 | null |
Stable Diffusion Now Optimized for AMD Radeon™ GPUs and Ryzen™ AI APUs — Stability AI
| 1 | 2025-04-18T06:48:08 |
https://stability.ai/news/stable-diffusion-now-optimized-for-amd-radeon-gpus
|
Primary-Wear-2460
|
stability.ai
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1ykre
| false | null |
t3_1k1ykre
|
/r/LocalLLaMA/comments/1k1ykre/stable_diffusion_now_optimized_for_amd_radeon/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'IkXeNjBdluDhw8S_2I3zEGHfDGKXWv8sPJN2Y6jnPfY', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/gHhwm7yG7wuXqGlE606RTVZfbDX9-mM6cPLepfpTuLQ.jpg?width=108&crop=smart&auto=webp&s=c08562bc27ea3070fe8d0302eec103ec4a9de2d8', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/gHhwm7yG7wuXqGlE606RTVZfbDX9-mM6cPLepfpTuLQ.jpg?width=216&crop=smart&auto=webp&s=a0a01cae33eaa90058e922ee224f933800f68082', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/gHhwm7yG7wuXqGlE606RTVZfbDX9-mM6cPLepfpTuLQ.jpg?width=320&crop=smart&auto=webp&s=36ec437494a12929b0349d032acf6bff108268a9', 'width': 320}, {'height': 640, 'url': 'https://external-preview.redd.it/gHhwm7yG7wuXqGlE606RTVZfbDX9-mM6cPLepfpTuLQ.jpg?width=640&crop=smart&auto=webp&s=597042a4c6f29c1d0879772ae9a3b8a06fabac7d', 'width': 640}, {'height': 960, 'url': 'https://external-preview.redd.it/gHhwm7yG7wuXqGlE606RTVZfbDX9-mM6cPLepfpTuLQ.jpg?width=960&crop=smart&auto=webp&s=f06e66a2f74c10a6fdd5da7689ffe5fe53a0da04', 'width': 960}, {'height': 1080, 'url': 'https://external-preview.redd.it/gHhwm7yG7wuXqGlE606RTVZfbDX9-mM6cPLepfpTuLQ.jpg?width=1080&crop=smart&auto=webp&s=ee677a347eb9b6bae482678bca615fddebb850c1', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/gHhwm7yG7wuXqGlE606RTVZfbDX9-mM6cPLepfpTuLQ.jpg?auto=webp&s=7cd4364774532dd9bde479583f33aa9856041448', 'width': 1080}, 'variants': {}}]}
|
||
Analyzing Technical Document Images with Janus-Pro 1B
| 1 |
I'm currently testing Janus-Pro for image analysis of technical documents, using the app from this GitHub repo: [`https://github.com/deepseek-ai/Janus`](https://github.com/deepseek-ai/Janus). I'm running it locally on a system with an Nvidia P4000 GPU (8GB VRAM), and I've switched the model from 7B to 1B to ensure it works on this hardware.
While it runs, the output tends to get cut off, and a lot of critical information is missing. Here's the image I'm using for input: [Janus Pro Plot and Graph](https://raw.githubusercontent.com/deepseek-ai/Janus/refs/heads/main/images/teaser_januspro.png)
Has anyone had better luck with Janus-Pro 1B? Were you able to get more complete or accurate outputs?
| 2025-04-18T06:52:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1yn4v/analyzing_technical_document_images_with_januspro/
|
kerkerby
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1yn4v
| false | null |
t3_1k1yn4v
|
/r/LocalLLaMA/comments/1k1yn4v/analyzing_technical_document_images_with_januspro/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'REu6cSXov90PNj67lx_19oEVywExCT5SE8dCCqUm8Kw', 'resolutions': [{'height': 47, 'url': 'https://external-preview.redd.it/7ZKZnoN8pJGulObWEYmKvylqUzjTj6bycuzfSUjxXgQ.png?width=108&crop=smart&auto=webp&s=839aaf59e78b46467446340f7e110cc837114b52', 'width': 108}, {'height': 95, 'url': 'https://external-preview.redd.it/7ZKZnoN8pJGulObWEYmKvylqUzjTj6bycuzfSUjxXgQ.png?width=216&crop=smart&auto=webp&s=88995c93e92fcde5c1fe2cefe334b54904f4b55a', 'width': 216}, {'height': 142, 'url': 'https://external-preview.redd.it/7ZKZnoN8pJGulObWEYmKvylqUzjTj6bycuzfSUjxXgQ.png?width=320&crop=smart&auto=webp&s=dbd7e543c35732ab5cd53b314fcf27df6cfed8b9', 'width': 320}, {'height': 284, 'url': 'https://external-preview.redd.it/7ZKZnoN8pJGulObWEYmKvylqUzjTj6bycuzfSUjxXgQ.png?width=640&crop=smart&auto=webp&s=471f546ba64d0a7386f8f67eb7552116cf9f7506', 'width': 640}], 'source': {'height': 407, 'url': 'https://external-preview.redd.it/7ZKZnoN8pJGulObWEYmKvylqUzjTj6bycuzfSUjxXgQ.png?auto=webp&s=83bd41f89d61ef70af2b19c55b63b4109316dcb8', 'width': 916}, 'variants': {}}]}
|
what is this
| 1 | 2025-04-18T06:59:24 |
Yareyaredaze880
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1yql6
| false | null |
t3_1k1yql6
|
/r/LocalLLaMA/comments/1k1yql6/what_is_this/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'lJk4PW_0yUjjQsXSE7h_wrTOd4QTDkpIlsD59jvWk84', 'resolutions': [{'height': 216, 'url': 'https://preview.redd.it/2vhlgw4okjve1.jpeg?width=108&crop=smart&auto=webp&s=8c873d44ae5220bc09b4b6cf1e7daf69421ffd00', 'width': 108}, {'height': 432, 'url': 'https://preview.redd.it/2vhlgw4okjve1.jpeg?width=216&crop=smart&auto=webp&s=87e647c7be70117e924118c055b28be52a7eb4b9', 'width': 216}, {'height': 640, 'url': 'https://preview.redd.it/2vhlgw4okjve1.jpeg?width=320&crop=smart&auto=webp&s=f2d6764d5e7b94a051596bf381fc9f038eac5518', 'width': 320}, {'height': 1280, 'url': 'https://preview.redd.it/2vhlgw4okjve1.jpeg?width=640&crop=smart&auto=webp&s=4bb3ef2b24ef761938b10b8a302524cf06520422', 'width': 640}, {'height': 1920, 'url': 'https://preview.redd.it/2vhlgw4okjve1.jpeg?width=960&crop=smart&auto=webp&s=79cb14d71eb08d13d42a49539b393b7d7ec1d44b', 'width': 960}, {'height': 2160, 'url': 'https://preview.redd.it/2vhlgw4okjve1.jpeg?width=1080&crop=smart&auto=webp&s=58b12793572b513cb0cd912b60d6a4aa79ab49a5', 'width': 1080}], 'source': {'height': 2400, 'url': 'https://preview.redd.it/2vhlgw4okjve1.jpeg?auto=webp&s=144fc873bc661e3c380b5268b9b642bc269aafd1', 'width': 1080}, 'variants': {}}]}
|
|||
New toy is here! A free general purpose, autonomous online AI agent!
| 2 |
I'm building a general purpose, autonomous online AI agent application ([kragent.ai](https://kragent.ai)) (yes, it's similar to Manus, but this project started long before Manus appeared). Currently, the core agent is already online, and I'm adding more available tools to it, including better MCP support. It's now open for free public beta testing. Hope everyone enjoys it.
Note, due to needing to control the costs of free operation, I've chosen the low-cost Qwen model. If you want to try the full-power effect, consider setting up your own Claude API key (don't bother trying OpenAI's models, they're particularly lazy when it comes to unofficial function calling mechanisms).
| 2025-04-18T07:39:36 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zas5/new_toy_is_here_a_free_general_purpose_autonomous/
|
Steven_Lu_137
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zas5
| false | null |
t3_1k1zas5
|
/r/LocalLLaMA/comments/1k1zas5/new_toy_is_here_a_free_general_purpose_autonomous/
| false | false |
self
| 2 | null |
What’s the most impressive local AI demo?
| 3 |
Imagine you have a group of nerdy and relatively tech-savvy friends ask you to show off some cool AI demo.
What would you show them to impress them and blow their mind?
| 2025-04-18T07:49:07 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zffp/whats_the_most_impressive_local_ai_demo/
|
lakySK
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zffp
| false | null |
t3_1k1zffp
|
/r/LocalLLaMA/comments/1k1zffp/whats_the_most_impressive_local_ai_demo/
| false | false |
self
| 3 | null |
Best vibe coding agent/tool?
| 0 |
I am really confused which tools have best performance. There are just too many of them like cursor, trae, windsurf, copilot, claude-code(cli), dozens of other agents on [swebench.com](http://swebench.com) leaderboards, and now open AI launched codex cli. It's not like the code quality is only dependent on the LLM model but also hugely affected by which environment/agent the model is used in. I have been using trae for a long time since it gave top models for free, but now I frequently run into rate limits. Also copilot is limited for free users even if I bring my own API which I am super pissed about. Is there any leaderboard which ranks all of them? Or anyone who tested all rigorously please shade some light.
| 2025-04-18T07:50:32 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zg5t/best_vibe_coding_agenttool/
|
Hv_V
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zg5t
| false | null |
t3_1k1zg5t
|
/r/LocalLLaMA/comments/1k1zg5t/best_vibe_coding_agenttool/
| false | false |
self
| 0 | null |
True local ChatGPT equivalent possible?
| 1 |
[removed]
| 2025-04-18T07:51:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zgqd/true_local_chatgpt_equivalent_possible/
|
Existing_Shoe_5594
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zgqd
| false | null |
t3_1k1zgqd
|
/r/LocalLLaMA/comments/1k1zgqd/true_local_chatgpt_equivalent_possible/
| false | false |
self
| 1 | null |
Just wondering about running llama 70b setup.
| 1 |
[removed]
| 2025-04-18T07:53:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zhj5/just_wondering_about_running_llama_70b_setup/
|
Appropriate-Tart-433
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zhj5
| false | null |
t3_1k1zhj5
|
/r/LocalLLaMA/comments/1k1zhj5/just_wondering_about_running_llama_70b_setup/
| false | false |
self
| 1 | null |
Just wondering about running Llama 70b.
| 1 |
[removed]
| 2025-04-18T07:54:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zi6p/just_wondering_about_running_llama_70b/
|
Appropriate-Tart-433
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zi6p
| false | null |
t3_1k1zi6p
|
/r/LocalLLaMA/comments/1k1zi6p/just_wondering_about_running_llama_70b/
| true | false |
spoiler
| 1 | null |
LLms about 70b suitable gpus combination recomendations?
| 1 |
[removed]
| 2025-04-18T07:56:29 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zj5t/llms_about_70b_suitable_gpus_combination/
|
Appropriate-Tart-433
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zj5t
| false | null |
t3_1k1zj5t
|
/r/LocalLLaMA/comments/1k1zj5t/llms_about_70b_suitable_gpus_combination/
| false | false |
self
| 1 | null |
LLms gpus suitable for 70b combination recomendations?
| 1 |
[removed]
| 2025-04-18T07:58:23 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zk1u/llms_gpus_suitable_for_70b_combination/
|
Appropriate-Tart-433
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zk1u
| false | null |
t3_1k1zk1u
|
/r/LocalLLaMA/comments/1k1zk1u/llms_gpus_suitable_for_70b_combination/
| false | false |
self
| 1 | null |
LLms gpus combination suitable for 70bs recomendations?
| 1 |
[removed]
| 2025-04-18T07:59:47 |
https://www.reddit.com/r/LocalLLaMA/comments/1k1zkq6/llms_gpus_combination_suitable_for_70bs/
|
Appropriate-Tart-433
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zkq6
| false | null |
t3_1k1zkq6
|
/r/LocalLLaMA/comments/1k1zkq6/llms_gpus_combination_suitable_for_70bs/
| false | false |
self
| 1 | null |
OpenAI API Codex connector
| 2 |
OpenAI had released their version of coding assistant as open source.
No big model library supports their Resources api yet, so they can’t work with it.
I wrote [a wrapper to make any OpenAI compatible library](https://github.com/OriNachum/openai-to-codex-wrapper), and verified it works (in the image you can see Mistral on Ollama)
It is still missing some features, but I would appreciate your support in stars, issues, suggestions and even pull requests if you are inclined for it.
I want to support the stateful features the other libraries don’t want to support and are needed for Codex (and more).
I verified it works in my main repo, in my [demo AI assistant that can hear, think and speak](https://github.com/OriNachum/autonomous-intelligence/tree/main/baby-tau) with the docker-compose-codex.yaml
Thank you for reading, and for your support if you are willing!
| 2025-04-18T08:22:26 |
Original_Finding2212
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zvym
| false | null |
t3_1k1zvym
|
/r/LocalLLaMA/comments/1k1zvym/openai_api_codex_connector/
| false | false | 2 |
{'enabled': True, 'images': [{'id': 'oVGHFRK0QTsbdtfmYKbruVRF_KMxdU7VEVt0uwfa4BM', 'resolutions': [{'height': 82, 'url': 'https://preview.redd.it/1o9uknihzjve1.jpeg?width=108&crop=smart&auto=webp&s=90a80822975df03af34379ed55c2514a46e6ffd9', 'width': 108}, {'height': 164, 'url': 'https://preview.redd.it/1o9uknihzjve1.jpeg?width=216&crop=smart&auto=webp&s=a65d47aa54e9aa1894c9d4e8d500f9d42707f7c5', 'width': 216}, {'height': 243, 'url': 'https://preview.redd.it/1o9uknihzjve1.jpeg?width=320&crop=smart&auto=webp&s=e463ed21ca7f1f586ce11c175d77d08bf585256f', 'width': 320}, {'height': 487, 'url': 'https://preview.redd.it/1o9uknihzjve1.jpeg?width=640&crop=smart&auto=webp&s=3b0a7d22bbf9047c3508aa582e5c2db28cf88dc5', 'width': 640}, {'height': 730, 'url': 'https://preview.redd.it/1o9uknihzjve1.jpeg?width=960&crop=smart&auto=webp&s=5b00be5742670af675f2ca15552dc04db4093b41', 'width': 960}], 'source': {'height': 819, 'url': 'https://preview.redd.it/1o9uknihzjve1.jpeg?auto=webp&s=c49d50c62839458965e0166289a7b7f58ce3cb14', 'width': 1076}, 'variants': {}}]}
|
||
Grok is cheaper & better than DeepSeek
| 0 |
Source : ArtificialAnlysis
| 2025-04-18T08:24:48 |
BidHot8598
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k1zx3j
| false | null |
t3_1k1zx3j
|
/r/LocalLLaMA/comments/1k1zx3j/grok_is_cheaper_better_than_deepseek/
| false | false | 0 |
{'enabled': True, 'images': [{'id': 'G4oZHzySJk989owvp7oqAciPdU2PjF9kSg8UVzK6pyw', 'resolutions': [{'height': 129, 'url': 'https://preview.redd.it/es97wnytzjve1.jpeg?width=108&crop=smart&auto=webp&s=315ebb93a8662b092c3de67cb91f3391925bff64', 'width': 108}, {'height': 259, 'url': 'https://preview.redd.it/es97wnytzjve1.jpeg?width=216&crop=smart&auto=webp&s=21ca5d7c89dfaa28da3b1a246b07d0fe80fde8eb', 'width': 216}, {'height': 384, 'url': 'https://preview.redd.it/es97wnytzjve1.jpeg?width=320&crop=smart&auto=webp&s=40525d54db21444bac40e6592f972194824b5c0b', 'width': 320}, {'height': 769, 'url': 'https://preview.redd.it/es97wnytzjve1.jpeg?width=640&crop=smart&auto=webp&s=777151dcb7e0e6ae384b2140bc88d07ece28d2f2', 'width': 640}], 'source': {'height': 915, 'url': 'https://preview.redd.it/es97wnytzjve1.jpeg?auto=webp&s=602a71cc9e369b5d934be2a82d080ecc5d5a1e69', 'width': 761}, 'variants': {}}]}
|
||
Best open source models ?
| 5 |
What are your top and best open source models ? And why ? no size restrictions .
| 2025-04-18T08:53:16 |
https://www.reddit.com/r/LocalLLaMA/comments/1k20ayu/best_open_source_models/
|
Basic-Pay-9535
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k20ayu
| false | null |
t3_1k20ayu
|
/r/LocalLLaMA/comments/1k20ayu/best_open_source_models/
| false | false |
self
| 5 | null |
st-r-awbe-rr-y
| 0 |
Help
| 2025-04-18T08:53:56 |
Infinite-Finance-515
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k20ba4
| false | null |
t3_1k20ba4
|
/r/LocalLLaMA/comments/1k20ba4/strawberry/
| false | false | 0 |
{'enabled': True, 'images': [{'id': 'M5VgcNmupiqstWzpuigzCaRumV2yT-a9QDmO-U9BDZw', 'resolutions': [{'height': 149, 'url': 'https://preview.redd.it/o89gxfp35kve1.jpeg?width=108&crop=smart&auto=webp&s=3b9b5490513105674559688157f1b337b45d9576', 'width': 108}, {'height': 299, 'url': 'https://preview.redd.it/o89gxfp35kve1.jpeg?width=216&crop=smart&auto=webp&s=1f0b8d4d0f6944a1b23db7a68a65dd5f98ae02c9', 'width': 216}, {'height': 444, 'url': 'https://preview.redd.it/o89gxfp35kve1.jpeg?width=320&crop=smart&auto=webp&s=9e006006ce3894b01683470b7c544d4fbb52d8af', 'width': 320}, {'height': 888, 'url': 'https://preview.redd.it/o89gxfp35kve1.jpeg?width=640&crop=smart&auto=webp&s=e1ee035d7a7c5c44ef566b8307ddd221ffb0f42f', 'width': 640}, {'height': 1332, 'url': 'https://preview.redd.it/o89gxfp35kve1.jpeg?width=960&crop=smart&auto=webp&s=af162a10d61ffa3b83c0cf368d344046e2f3f15b', 'width': 960}, {'height': 1498, 'url': 'https://preview.redd.it/o89gxfp35kve1.jpeg?width=1080&crop=smart&auto=webp&s=89c5e4717bcb7d24b1c8c47dc18610b02caaf4dc', 'width': 1080}], 'source': {'height': 1561, 'url': 'https://preview.redd.it/o89gxfp35kve1.jpeg?auto=webp&s=adc1ba4bcd874fa575e5ad43153a89084e68979f', 'width': 1125}, 'variants': {}}]}
|
||
Is there a small tool-calling LLM?
| 15 |
So basically i want to do an LLM game engine that resolves missing stuff via an llm. For that i need an LLM which complies with tool calling and actually calls tools whenever there's an opportunity. Is there such an LLM, that's small enough to not boil my room? Ideally a 7B one, it just needs to follow instructions it gets from tool calls.
| 2025-04-18T10:47:56 |
https://www.reddit.com/r/LocalLLaMA/comments/1k21yub/is_there_a_small_toolcalling_llm/
|
ashleigh_dashie
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k21yub
| false | null |
t3_1k21yub
|
/r/LocalLLaMA/comments/1k21yub/is_there_a_small_toolcalling_llm/
| false | false |
self
| 15 | null |
Good news: 5090s now in stock in my local market. Bad news: cheapest is $3,550
| 49 |
Now I wonder if I sshould have just bought the 2nd hand 3090s that were on sale for $700.
| 2025-04-18T11:13:40 |
https://www.reddit.com/r/LocalLLaMA/comments/1k22e41/good_news_5090s_now_in_stock_in_my_local_market/
|
DeltaSqueezer
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k22e41
| false | null |
t3_1k22e41
|
/r/LocalLLaMA/comments/1k22e41/good_news_5090s_now_in_stock_in_my_local_market/
| false | false |
self
| 49 | null |
OpenAI naming is so confusing they need to include explanations inside Codex CLI system prompt
| 17 |
I was going through Codex CLI system prompt and found this gem.
As a reminder OpenAI released Codex LLM tuned for coding couple of years back.
Here’s the excerpt:
“The Codex CLI is open-sourced. Don't confuse yourself with the old Codex language model built by OpenAI many moons ago (this is understandably top of mind for you!). Within this context, Codex refers to the open-source agentic coding interface.”
| 2025-04-18T11:20:26 |
https://github.com/openai/codex/blob/main/codex-cli/src/utils/agent/agent-loop.ts
|
fixtwin
|
github.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k22i44
| false | null |
t3_1k22i44
|
/r/LocalLLaMA/comments/1k22i44/openai_naming_is_so_confusing_they_need_to/
| false | false | 17 |
{'enabled': False, 'images': [{'id': 'DHbmFtBH5i-qv266Zeh2R316QgFWX3u4FD4L1BwZtio', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/GzGxq7sjEIOdqa06p0Tb5ISsD5qYfTkIq7FrR23kigg.jpg?width=108&crop=smart&auto=webp&s=8f204065a39624c33e82bc63cb8fffce7b281393', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/GzGxq7sjEIOdqa06p0Tb5ISsD5qYfTkIq7FrR23kigg.jpg?width=216&crop=smart&auto=webp&s=b09ea16ef924fc3edd24be92f194eeba6c2476ba', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/GzGxq7sjEIOdqa06p0Tb5ISsD5qYfTkIq7FrR23kigg.jpg?width=320&crop=smart&auto=webp&s=68797f4f7815b7e6f7ec1716ab96c6c2da0c6769', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/GzGxq7sjEIOdqa06p0Tb5ISsD5qYfTkIq7FrR23kigg.jpg?width=640&crop=smart&auto=webp&s=dfa0db6f6bef47eb4fc9a05e50dd4119e49a45df', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/GzGxq7sjEIOdqa06p0Tb5ISsD5qYfTkIq7FrR23kigg.jpg?width=960&crop=smart&auto=webp&s=e5aa440e535d77e216ba813325979549e9d85de6', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/GzGxq7sjEIOdqa06p0Tb5ISsD5qYfTkIq7FrR23kigg.jpg?width=1080&crop=smart&auto=webp&s=d460e3769d05f31bfa7047ea732a3b4868da06af', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/GzGxq7sjEIOdqa06p0Tb5ISsD5qYfTkIq7FrR23kigg.jpg?auto=webp&s=59be9ec96100875714833fcfa9fad49155b912f3', 'width': 1200}, 'variants': {}}]}
|
|
FULL LEAKED Replit Agent System Prompts and Tools
| 69 |
(Latest system prompt: 18/04/2025)
I managed to get full official Replit Agent system prompts, including its tools (JSON). Over 400 lines.
You can check it out at: [https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools](https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools)
| 2025-04-18T11:26:40 |
https://www.reddit.com/r/LocalLLaMA/comments/1k22lyx/full_leaked_replit_agent_system_prompts_and_tools/
|
Independent-Box-898
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k22lyx
| false | null |
t3_1k22lyx
|
/r/LocalLLaMA/comments/1k22lyx/full_leaked_replit_agent_system_prompts_and_tools/
| false | false |
self
| 69 | null |
How to get started with the new Dream 7B diffusion reasoning model
| 1 |
[removed]
| 2025-04-18T11:52:10 |
https://www.youtube.com/watch?v=xyLuihgD9BA
|
ai-local
|
youtube.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k231vr
| false |
{'oembed': {'author_name': "Brian's Linux and AIX Videos", 'author_url': 'https://www.youtube.com/@ixbrian0', 'height': 200, 'html': '<iframe width="356" height="200" src="https://www.youtube.com/embed/xyLuihgD9BA?feature=oembed&enablejsapi=1" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen title="How to get started with the Dream 7B diffusion reasoning model"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/xyLuihgD9BA/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'How to get started with the Dream 7B diffusion reasoning model', 'type': 'video', 'version': '1.0', 'width': 356}, 'type': 'youtube.com'}
|
t3_1k231vr
|
/r/LocalLLaMA/comments/1k231vr/how_to_get_started_with_the_new_dream_7b/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'pA0-5Cn1x5Xd57rNrpcD5PCXpaIBGse8rwqFaJgyd6k', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/VmG8y_6yIHgSOZvKQdnqMhlkmXZtG5GcwOTyIXxA4rc.jpg?width=108&crop=smart&auto=webp&s=ce06c3a733e6f6a39ca40277bc563e05663f1569', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/VmG8y_6yIHgSOZvKQdnqMhlkmXZtG5GcwOTyIXxA4rc.jpg?width=216&crop=smart&auto=webp&s=e17d850133eab0488379f8f3e8b96c8ae08c585c', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/VmG8y_6yIHgSOZvKQdnqMhlkmXZtG5GcwOTyIXxA4rc.jpg?width=320&crop=smart&auto=webp&s=1623aacf898adb02968e5de20711f3755faf4bb2', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/VmG8y_6yIHgSOZvKQdnqMhlkmXZtG5GcwOTyIXxA4rc.jpg?auto=webp&s=9c0f92332d0db33c58693d63dee8242874df0429', 'width': 480}, 'variants': {}}]}
|
|
Does anyone else feel guilty using big models for tiny tasks?
| 18 |
I don't know if anyone else feels this way, but sometimes when I use a huge model for something super simple, I feel bad, like I'm wasting resources or something.
It feels like these LLMs are way too powerful for little tasks, and I shouldn't be wasting their "time" (even though I know it's not alive lol) or the computational resources.
Because of that, I set up Gemma 3 locally and now I use it for all my tiny tasks.
I can't fully explain why I feel like this — it's not really logical — but it's there.
Does anyone else feel the same way?
| 2025-04-18T12:02:06 |
https://www.reddit.com/r/LocalLLaMA/comments/1k238lw/does_anyone_else_feel_guilty_using_big_models_for/
|
RightCup5772
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k238lw
| false | null |
t3_1k238lw
|
/r/LocalLLaMA/comments/1k238lw/does_anyone_else_feel_guilty_using_big_models_for/
| false | false |
self
| 18 | null |
Fuzzy quant scaling for dynamic reasoning steps.
| 0 |
Hear me out, and you geniuses may understand.
So as part of reasoning it's valuable to step back from the immediate issue and be a little more broad and encompassing.
What would be the effect of adding a controlled and intelligently scaled amount of noise to the weights during inference?
Maybe just inside specific trigger tags you fudge the math a little to produce a slightly noisy gradient?
Could this gentle fuzz lead to better reasoning divergence while maintaining coherence and staying near topic?
It's important to note that I don't mean consistent changes, I mean dynamic and optional fuzzy weights per token with some type of controls for activation and curve.
Do something fancy with the context data to optimize per token or something. My expectation is someone smarter than me will know more exactly about how the math works.
All I know for sure about how the math shakes out is if you shoot some marbles onto 10B semi directional pinball bumpers and collect the marbles that escape there will be areas where lots of marbles stop together and the decoder layer turns that into numbers that relate to words or groups of words and their probability:
[ [306627" cow",0.7673],[100837" chocolate milk", 0.19631]]
The prompt controls how and where you shoot the marbles, there are 128k or 32k holes around the perimeter per model. One for each vocabulary token.
Just a wee noise to simulate the jostle and consistent yet unpredictable real pinball experience and shake the really certain models up a bit that isn't based around random sampling the final outs. Might be something to gain. Might be nonsense. I can't decide if it's gibberish or if it might help in reasoning and review on some models and tasks.
Anyway, cool chat. I'm probably ignorant of a large barrier to implementation and speed would lilely be significantly degraded. I don't have time or quiet to sink into the code. It's on you guys.
Thanks for reading.
| 2025-04-18T12:05:59 |
https://www.reddit.com/r/LocalLLaMA/comments/1k23b3a/fuzzy_quant_scaling_for_dynamic_reasoning_steps/
|
aseichter2007
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k23b3a
| false | null |
t3_1k23b3a
|
/r/LocalLLaMA/comments/1k23b3a/fuzzy_quant_scaling_for_dynamic_reasoning_steps/
| false | false |
self
| 0 | null |
RTX 3090 24GB gotten! Initial panic but all good - now some AI practic🥳
| 1 |
[removed]
| 2025-04-18T12:11:34 |
https://www.reddit.com/gallery/1k23eto
|
hsangh
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k23eto
| false | null |
t3_1k23eto
|
/r/LocalLLaMA/comments/1k23eto/rtx_3090_24gb_gotten_initial_panic_but_all_good/
| false | false | 1 | null |
|
Do any of the gpt-4.1 or o3/o4 models in the API support search?
| 0 |
I'm not sure if I'm doing something wrong, but querying gpt-4.1 and o4-mini through the openai API endpoint doesn't seem to be able to search up anything online (the way the gpt-4o-search-preview-2025-03-11 model could).
Do these models support search by themselves, or do we need to have it call a tool?
| 2025-04-18T12:19:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1k23ka1/do_any_of_the_gpt41_or_o3o4_models_in_the_api/
|
jaxchang
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k23ka1
| false | null |
t3_1k23ka1
|
/r/LocalLLaMA/comments/1k23ka1/do_any_of_the_gpt41_or_o3o4_models_in_the_api/
| false | false |
self
| 0 | null |
Awesome Multimodal Search
| 1 | 2025-04-18T12:51:35 |
https://github.com/mixpeek/awesome-multimodal-search/tree/main
|
Feeling_Mushroom3175
|
github.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k246fl
| false | null |
t3_1k246fl
|
/r/LocalLLaMA/comments/1k246fl/awesome_multimodal_search/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'Ytl6kOE1ZqMaq3XsQrD1Fc6sPgquuYxRZRKPaos5zIo', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/LOFclA20IL7Fe3sJelaVN97hCHjrTKlGmcvvV42ydoI.jpg?width=108&crop=smart&auto=webp&s=a40cc21dadcc04f081d74c3d7fde4668482b31b9', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/LOFclA20IL7Fe3sJelaVN97hCHjrTKlGmcvvV42ydoI.jpg?width=216&crop=smart&auto=webp&s=51ba1842588ca38a822a2a07a754fed73bdf973c', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/LOFclA20IL7Fe3sJelaVN97hCHjrTKlGmcvvV42ydoI.jpg?width=320&crop=smart&auto=webp&s=c686bc90b7d9dd93d4d4344409c5074b91df8207', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/LOFclA20IL7Fe3sJelaVN97hCHjrTKlGmcvvV42ydoI.jpg?width=640&crop=smart&auto=webp&s=ed3816a766fa98f43a2068d66d9fd141927936f7', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/LOFclA20IL7Fe3sJelaVN97hCHjrTKlGmcvvV42ydoI.jpg?width=960&crop=smart&auto=webp&s=de6cf5a2223d733b2e71d9eb992ecb3c990a5d84', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/LOFclA20IL7Fe3sJelaVN97hCHjrTKlGmcvvV42ydoI.jpg?width=1080&crop=smart&auto=webp&s=ff40b5a58f6a13a2403c32f3e95ceecf22d1211b', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/LOFclA20IL7Fe3sJelaVN97hCHjrTKlGmcvvV42ydoI.jpg?auto=webp&s=42d3f0fb91055b194ed9a3810a7c9cd74cb03ca2', 'width': 1200}, 'variants': {}}]}
|
||
Google’s Agent2Agent (A2A) Explained
| 8 |
Hey everyone,
Just published a new \*FREE\* blog post on Agent-to-Agent (A2A) – Google’s new framework letting AI systems collaborate like human teammates rather than working in isolation.
In this post, I explain:
\- Why specialized AI agents need to talk to each other
\- How A2A compares to MCP and why they're complementary
\- The essentials of A2A
I've kept it accessible with real-world examples like planning a birthday party. This approach represents a fundamental shift where we'll delegate to teams of AI agents working together rather than juggling specialized tools ourselves.
Link to the full blog post:
[https://open.substack.com/pub/diamantai/p/googles-agent2agent-a2a-explained?r=336pe4&utm\_campaign=post&utm\_medium=web&showWelcomeOnShare=false](https://open.substack.com/pub/diamantai/p/googles-agent2agent-a2a-explained?r=336pe4&utm_campaign=post&utm_medium=web&showWelcomeOnShare=false)
| 2025-04-18T12:56:36 |
https://www.reddit.com/r/LocalLLaMA/comments/1k249wy/googles_agent2agent_a2a_explained/
|
Nir777
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k249wy
| false | null |
t3_1k249wy
|
/r/LocalLLaMA/comments/1k249wy/googles_agent2agent_a2a_explained/
| false | false |
self
| 8 |
{'enabled': False, 'images': [{'id': 'HHQrUWrKmhwThNg5GCEs9lUuU6WlZCoWcmzSuCgWtl0', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/7QXEOnqrCYHOKK7JNIbEmqU70dUZsY4OiMRh_DIUbTI.jpg?width=108&crop=smart&auto=webp&s=be34a9fbe8beb066b7df53b93ac4ce81b3508bff', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/7QXEOnqrCYHOKK7JNIbEmqU70dUZsY4OiMRh_DIUbTI.jpg?width=216&crop=smart&auto=webp&s=648a49adf32501e9c4294b391ec9e384c4b27483', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/7QXEOnqrCYHOKK7JNIbEmqU70dUZsY4OiMRh_DIUbTI.jpg?width=320&crop=smart&auto=webp&s=c3035804d990ee7a47cb5de2cb2f9837ae205bcc', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/7QXEOnqrCYHOKK7JNIbEmqU70dUZsY4OiMRh_DIUbTI.jpg?width=640&crop=smart&auto=webp&s=7edcc2ffa0b0ebe465fb68ab5e28677c7337bf0c', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/7QXEOnqrCYHOKK7JNIbEmqU70dUZsY4OiMRh_DIUbTI.jpg?width=960&crop=smart&auto=webp&s=312795e868044a8d85e023af4d6e37117f0fcc1a', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/7QXEOnqrCYHOKK7JNIbEmqU70dUZsY4OiMRh_DIUbTI.jpg?width=1080&crop=smart&auto=webp&s=d15426445d9355366fd001aa6497afbcb7360916', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/7QXEOnqrCYHOKK7JNIbEmqU70dUZsY4OiMRh_DIUbTI.jpg?auto=webp&s=02e43352ed6fc6cd3ac312638c73bdb70d1df9d3', 'width': 1200}, 'variants': {}}]}
|
Is there a tutorial on how to write a CYOA story?
| 1 |
[removed]
| 2025-04-18T12:57:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1k24alh/is_there_a_tutorial_on_how_to_write_a_cyoa_story/
|
BlacksmithRadiant322
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k24alh
| false | null |
t3_1k24alh
|
/r/LocalLLaMA/comments/1k24alh/is_there_a_tutorial_on_how_to_write_a_cyoa_story/
| false | false |
self
| 1 | null |
I know these aren't local but Veo2 is just insane, I hope we get this level of quality locally soon.
| 0 |
https://imgur.com/0i40ybQ
https://imgur.com/HB4TrMA
| 2025-04-18T13:10:00 |
https://www.reddit.com/r/LocalLLaMA/comments/1k24jz3/i_know_these_arent_local_but_veo2_is_just_insane/
|
Sea_Sympathy_495
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k24jz3
| false | null |
t3_1k24jz3
|
/r/LocalLLaMA/comments/1k24jz3/i_know_these_arent_local_but_veo2_is_just_insane/
| false | false |
self
| 0 |
{'enabled': False, 'images': [{'id': '-k9Y1ztr1uI0J93nErMFTquhyBDMVWHmDvyyXMWEybw', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/wEhelBhsvvAWRHZB_HFVDxJqrHitHhxugr2aKVG595Y.jpg?width=108&crop=smart&auto=webp&s=57e93892a5679b3a9a20003c2cba7a31f6ee46f1', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/wEhelBhsvvAWRHZB_HFVDxJqrHitHhxugr2aKVG595Y.jpg?width=216&crop=smart&auto=webp&s=19764701ebc6e4dd189c8c2de288fa5145fdf918', 'width': 216}, {'height': 179, 'url': 'https://external-preview.redd.it/wEhelBhsvvAWRHZB_HFVDxJqrHitHhxugr2aKVG595Y.jpg?width=320&crop=smart&auto=webp&s=74a74b82414061d80a91f83ca96adebe2ba38610', 'width': 320}, {'height': 359, 'url': 'https://external-preview.redd.it/wEhelBhsvvAWRHZB_HFVDxJqrHitHhxugr2aKVG595Y.jpg?width=640&crop=smart&auto=webp&s=34413dd84410481b2611b07986ca12125ef9e32b', 'width': 640}], 'source': {'height': 480, 'url': 'https://external-preview.redd.it/wEhelBhsvvAWRHZB_HFVDxJqrHitHhxugr2aKVG595Y.jpg?auto=webp&s=ee0f7ea8b8d8d0a2288735ac632dca9642308d3f', 'width': 854}, 'variants': {}}]}
|
It’s merge time! - What are your top tips for using MergeKit?
| 3 |
I recently posted about the [release of Minueza 2](https://www.reddit.com/r/LocalLLaMA/comments/1jt3vvq/minueza296m_a_foundation_bilingual_textgeneration/), and today I completed the upload of [10 different fine-tunes](https://huggingface.co/collections/Felladrin/minueza-2-96m-67f2ca2b07a3da4558d5c3f6), ready for merging experiments!
I’m aware of simple merges, as listed in the [examples of mergekit](https://github.com/arcee-ai/mergekit/tree/main/examples), and complex ones, as described in the [model card of Qwen2.5-14B-Vimarckoso-v3](https://huggingface.co/sometimesanotion/Qwen2.5-14B-Vimarckoso-v3), but I’d love to learn about other techniques that you’ve been using.
| 2025-04-18T13:12:22 |
https://www.reddit.com/r/LocalLLaMA/comments/1k24lpi/its_merge_time_what_are_your_top_tips_for_using/
|
Felladrin
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k24lpi
| false | null |
t3_1k24lpi
|
/r/LocalLLaMA/comments/1k24lpi/its_merge_time_what_are_your_top_tips_for_using/
| false | false |
self
| 3 |
{'enabled': False, 'images': [{'id': '0CLwgDbHKWpqtgtG55PGvmG-ZU6b1YME02BJV4MksQY', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/jaXYFGaAf2FAsz7LbBPArTI_Q5pIKGEKWI2XZLpzS0M.jpg?width=108&crop=smart&auto=webp&s=df33122d7aaac29c5778d055252c1e4e7c33664a', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/jaXYFGaAf2FAsz7LbBPArTI_Q5pIKGEKWI2XZLpzS0M.jpg?width=216&crop=smart&auto=webp&s=032d943e1d33a055b20725a674925a46e232517a', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/jaXYFGaAf2FAsz7LbBPArTI_Q5pIKGEKWI2XZLpzS0M.jpg?width=320&crop=smart&auto=webp&s=9898e804e91b8d185b9207c2506c02e375dfce8a', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/jaXYFGaAf2FAsz7LbBPArTI_Q5pIKGEKWI2XZLpzS0M.jpg?width=640&crop=smart&auto=webp&s=8f9d4c4ca769f70f6c10ebc40ae88af3ce904ce1', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/jaXYFGaAf2FAsz7LbBPArTI_Q5pIKGEKWI2XZLpzS0M.jpg?width=960&crop=smart&auto=webp&s=954a9ce06458e0baa46a626aa2dc00d97823291b', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/jaXYFGaAf2FAsz7LbBPArTI_Q5pIKGEKWI2XZLpzS0M.jpg?width=1080&crop=smart&auto=webp&s=0e49922fc5e9e882a3970ee8286fda69995f08df', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/jaXYFGaAf2FAsz7LbBPArTI_Q5pIKGEKWI2XZLpzS0M.jpg?auto=webp&s=e871f9d0d99f8805245ae829f1d40a2d6fd15246', 'width': 1200}, 'variants': {}}]}
|
Can I run any LLM on my potato laptop?
| 3 |
I have i5 a laptop with 8gm, is it possible to run any model on it ? if so.. then which one?
| 2025-04-18T13:16:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1k24osy/can_i_run_any_llm_on_my_potato_laptop/
|
Both-Drama-8561
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k24osy
| false | null |
t3_1k24osy
|
/r/LocalLLaMA/comments/1k24osy/can_i_run_any_llm_on_my_potato_laptop/
| false | false |
self
| 3 | null |
Judging Embeddings
| 0 |
When considering embeddings, it can be a helpful qualitative tool to check the top-k most similar results in a neighborhood of your query samples. This can help you find clear themes and patterns to explain how your model organizes the data.
But its a slow, subjective technique so I'm thinking about applying VLM-as-a-Judge, prompting AI to identify themes explaining the cluster and scoring it quantitatively.
Zero-shot without much experimenting with the prompt for a generic model but the technique looks promising. I tried this idea on my custom theatrical poster embeddings, made before CLIP was open-sourced.
Can Judging Embeddings help make your AI experiments more explainable?
| 2025-04-18T13:16:47 |
https://www.reddit.com/gallery/1k24p1z
|
remyxai
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k24p1z
| false | null |
t3_1k24p1z
|
/r/LocalLLaMA/comments/1k24p1z/judging_embeddings/
| false | false | 0 | null |
|
Gemma 3 QAT launch with MLX, llama.cpp, Ollama, LM Studio, and Hugging Face
| 204 |
Hi!
Some weeks ago we released GGUFs corresponding to the QAT checkpoints of Gemma 3. Thanks to QAT, the model is able to preserve similar quality as `bfloat16` while significantly reducing the memory requirements to load the model. That is, QAT is an additional fine-tuning that makes the model more rigorous to quantization.
As we only released the GGUFs, we got feedback that it would be great to have the unquantized QAT-based checkpoints to allow people to quantize for their own tools. So...we did it! Today we're releasing the unquantized QAT-based checkpoints. The models preserve quality better than naive quantization.
**We also collaborated with Prince (from MLX), llama.cpp, Ollama, LM Studio, and Hugging Face to make sure you can use the models in all your favorite tools!**
* Blog post : [https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/](https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/)
* Unquantized checkpoints: [https://huggingface.co/collections/google/gemma-3-qat-67ee61ccacbf2be4195c265b](https://huggingface.co/collections/google/gemma-3-qat-67ee61ccacbf2be4195c265b)
* Ollama: [https://ollama.com/library/gemma3](https://ollama.com/library/gemma3) (try ollama run gemma3:12b-it-qat)
* LM Studio: [https://lmstudio.ai/model/gemma-3-12b-it-qat](https://lmstudio.ai/model/gemma-3-12b-it-qat)
* MLX: [https://huggingface.co/collections/mlx-community/gemma-3-qat-68002674cd5afc6f9022a0ae](https://huggingface.co/collections/mlx-community/gemma-3-qat-68002674cd5afc6f9022a0ae)
* llama.cpp: [https://huggingface.co/collections/google/gemma-3-qat-67ee61ccacbf2be4195c265b](https://huggingface.co/collections/google/gemma-3-qat-67ee61ccacbf2be4195c265b)
Enjoy!
| 2025-04-18T13:31:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1k250fu/gemma_3_qat_launch_with_mlx_llamacpp_ollama_lm/
|
hackerllama
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k250fu
| false | null |
t3_1k250fu
|
/r/LocalLLaMA/comments/1k250fu/gemma_3_qat_launch_with_mlx_llamacpp_ollama_lm/
| false | false |
self
| 204 |
{'enabled': False, 'images': [{'id': 'UPh_4CgafUqTh9ZB3bC0-0Msh-CF5QgkiP-Ex1y8M_I', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=108&crop=smart&auto=webp&s=bf80d9b78a582598ddaf46ebb198ba14da0dfee1', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=216&crop=smart&auto=webp&s=1ad09b95d0279438bd66d1d418f3f9e0b207e8d8', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=320&crop=smart&auto=webp&s=75848e136c8a8aa2ea0df8ba00019ebccfebb3fe', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=640&crop=smart&auto=webp&s=ed6a861b423ef5ef481e863b5c6947b3cef14c0c', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=960&crop=smart&auto=webp&s=7486c5a54c0c3728faa8358805c8f52cd7e039fd', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=1080&crop=smart&auto=webp&s=c783ae4f78dd78266208a25dc198c7d56dcb9de7', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?auto=webp&s=b9fa62cfd071dc2a391de1c697f0bfbb56d04afa', 'width': 1200}, 'variants': {}}]}
|
New QAT-optimized int4 Gemma 3 models by Google, slash VRAM needs (54GB -> 14.1GB) while maintaining quality.
| 352 | 2025-04-18T13:32:01 |
https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/?linkId=14034718
|
Sea_Sympathy_495
|
developers.googleblog.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1k250r6
| false | null |
t3_1k250r6
|
/r/LocalLLaMA/comments/1k250r6/new_qatoptimized_int4_gemma_3_models_by_google/
| false | false | 352 |
{'enabled': False, 'images': [{'id': 'UPh_4CgafUqTh9ZB3bC0-0Msh-CF5QgkiP-Ex1y8M_I', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=108&crop=smart&auto=webp&s=bf80d9b78a582598ddaf46ebb198ba14da0dfee1', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=216&crop=smart&auto=webp&s=1ad09b95d0279438bd66d1d418f3f9e0b207e8d8', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=320&crop=smart&auto=webp&s=75848e136c8a8aa2ea0df8ba00019ebccfebb3fe', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=640&crop=smart&auto=webp&s=ed6a861b423ef5ef481e863b5c6947b3cef14c0c', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=960&crop=smart&auto=webp&s=7486c5a54c0c3728faa8358805c8f52cd7e039fd', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?width=1080&crop=smart&auto=webp&s=c783ae4f78dd78266208a25dc198c7d56dcb9de7', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/5lq32BTIzHqmPYcHvNrCp8JMhag9gsSSkR3cQgoYZBU.jpg?auto=webp&s=b9fa62cfd071dc2a391de1c697f0bfbb56d04afa', 'width': 1200}, 'variants': {}}]}
|
||
Google new Gemma 3 QAT - optimized int4 models slash VRAM needs (54GB -> 14.1GB) while maintaining quality
| 1 |
[removed]
| 2025-04-18T13:37:23 |
Nunki08
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k254t2
| false | null |
t3_1k254t2
|
/r/LocalLLaMA/comments/1k254t2/google_new_gemma_3_qat_optimized_int4_models/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'dZkxbbLwH-IffOEaILftM1VFyPW3z_BszqFCxXS7Tnc', 'resolutions': [{'height': 75, 'url': 'https://preview.redd.it/vx4tvzerilve1.jpeg?width=108&crop=smart&auto=webp&s=9d6c8e7230d997d524d80bb9ed325e8e17271ff2', 'width': 108}, {'height': 150, 'url': 'https://preview.redd.it/vx4tvzerilve1.jpeg?width=216&crop=smart&auto=webp&s=9ee381be40748a1dfca9819277d599b56e7a04a4', 'width': 216}, {'height': 223, 'url': 'https://preview.redd.it/vx4tvzerilve1.jpeg?width=320&crop=smart&auto=webp&s=de9d852234e343655c5c2702129d32b4bd7e0f2f', 'width': 320}, {'height': 446, 'url': 'https://preview.redd.it/vx4tvzerilve1.jpeg?width=640&crop=smart&auto=webp&s=531f96adaf06b8da3bbfec267400d57b028e62c5', 'width': 640}, {'height': 670, 'url': 'https://preview.redd.it/vx4tvzerilve1.jpeg?width=960&crop=smart&auto=webp&s=48e3ac5c26b8d5258662e69bdc014a2ad4a1bfc2', 'width': 960}, {'height': 754, 'url': 'https://preview.redd.it/vx4tvzerilve1.jpeg?width=1080&crop=smart&auto=webp&s=7417140a93b5ecc9deed75720f929540097d8d41', 'width': 1080}], 'source': {'height': 1410, 'url': 'https://preview.redd.it/vx4tvzerilve1.jpeg?auto=webp&s=83321526bc952ad2e0c19f1790ff995cf3852fe3', 'width': 2019}, 'variants': {}}]}
|
||
Google QAT - optimized int4 Gemma 3 slash VRAM needs (54GB -> 14.1GB) while maintaining quality - llama.cpp, lmstudio, MLX, ollama
| 701 | 2025-04-18T13:41:47 |
Nunki08
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k25876
| false | null |
t3_1k25876
|
/r/LocalLLaMA/comments/1k25876/google_qat_optimized_int4_gemma_3_slash_vram/
| false | false | 701 |
{'enabled': True, 'images': [{'id': 'WpG3vdmaKqlrkdv92vZkx1C4XsRcB60nvCuzf1WOIVE', 'resolutions': [{'height': 75, 'url': 'https://preview.redd.it/23ut7jd3klve1.jpeg?width=108&crop=smart&auto=webp&s=d24a72cfeba57d1329030822c53b165171d81cc7', 'width': 108}, {'height': 150, 'url': 'https://preview.redd.it/23ut7jd3klve1.jpeg?width=216&crop=smart&auto=webp&s=f4249818f6ce5f595fc6fd8bf7119aa05e1249ea', 'width': 216}, {'height': 223, 'url': 'https://preview.redd.it/23ut7jd3klve1.jpeg?width=320&crop=smart&auto=webp&s=cbdff285fa824e92fc9f36614da8efdb2852072f', 'width': 320}, {'height': 446, 'url': 'https://preview.redd.it/23ut7jd3klve1.jpeg?width=640&crop=smart&auto=webp&s=f940165ab5ba660103d9f5f61872b1dc70698cbb', 'width': 640}, {'height': 670, 'url': 'https://preview.redd.it/23ut7jd3klve1.jpeg?width=960&crop=smart&auto=webp&s=f0bd9f243f94109afab58822dcdd571b442e61ef', 'width': 960}, {'height': 754, 'url': 'https://preview.redd.it/23ut7jd3klve1.jpeg?width=1080&crop=smart&auto=webp&s=51670d55165b6bdea049b26ad1e0e9d1cc0c630b', 'width': 1080}], 'source': {'height': 1410, 'url': 'https://preview.redd.it/23ut7jd3klve1.jpeg?auto=webp&s=e3e2f50b8702624aa44bca7507d152bb8230ab03', 'width': 2019}, 'variants': {}}]}
|
|||
How to Improve Search Accuracy in a Retrieval System?
| 5 |
Hey everyone,
I’m working on a small RAG setup that lets users search vehicle‑event image captions (e.g., “driver wearing red”). I’m using Milvus’s hybrid search with BAAI/bge‑m3 to generate both dense and sparse embeddings, but I keep running into accuracy issues. For example, it often returns captions about “red vehicle” where the driver is wearing a completely different color—even with very high scores. I also tried adding a reranker (BAAI/bge‑reranker‑v2‑m3), but noticed no improvement.
What I need help with:
* How can I get more precise results for my use-case?
* How do you evaluate search accuracy in this context? Is there an existing framework or set of metrics I can use?
I’d really appreciate any advice or examples. Thanks!
| 2025-04-18T13:45:30 |
https://www.reddit.com/r/LocalLLaMA/comments/1k25b0v/how_to_improve_search_accuracy_in_a_retrieval/
|
Traditional_Tap1708
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k25b0v
| false | null |
t3_1k25b0v
|
/r/LocalLLaMA/comments/1k25b0v/how_to_improve_search_accuracy_in_a_retrieval/
| false | false |
self
| 5 | null |
Privacy of local, Power of cloud?
| 1 |
[removed]
| 2025-04-18T13:59:50 |
https://www.reddit.com/r/LocalLLaMA/comments/1k25mdw/privacy_of_local_power_of_cloud/
|
judethedude
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k25mdw
| false | null |
t3_1k25mdw
|
/r/LocalLLaMA/comments/1k25mdw/privacy_of_local_power_of_cloud/
| false | false |
self
| 1 | null |
GPT 4.1 is a game changer
| 0 |
I've been working on a few multilingual text forecasting projects for a while now. I have been a staunch user of Llama 3.1 8B just based on how well it does after fine-tuning on my (pretty difficult) forecasting benchmarks. My ROC-AUCs have hovered close to 0.8 for the best models. Llama 3.1 8B performed comparably to GPT-4o *and* GPT-4o-mini, so I had written off my particular use case as too difficult for bigger models.
I fine-tuned GPT 4.1 earlier today and achieved an ROC-AUC of 0.94. This is a game changer; it essentially "solves" my particular class of problems. I have to get rid of an entire Llama-based reinforcement learning pipeline I literally just built over the past month.
This is just a PSA if any of you are considering whether it's worth fine-tuning GPT 4.1. It cost me a few $100s for both fine-tuning and inference. My H100 GPU cost $25,000 and I'm now regretting the purchase. I didn't believe in model scaling laws, now I do.
| 2025-04-18T14:07:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1k25suh/gpt_41_is_a_game_changer/
|
entsnack
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k25suh
| false | null |
t3_1k25suh
|
/r/LocalLLaMA/comments/1k25suh/gpt_41_is_a_game_changer/
| false | false |
self
| 0 | null |
Multi-Node Cluster Deployment of Qwen Series Models with SGLang
| 4 |
### Objective
While `Ollama` offers convenience, high concurrency is sometimes more crucial. This article demonstrates how to deploy `SGLang` on two computers (dual nodes) to run the `Qwen2.5-7B-Instruct` model, maximizing local resource utilization. Additional nodes can be added if available.
### Hardware Requirements
- Node 0: IP `192.168.0.12`, 1 NVIDIA GPU
- Node 1: IP `192.168.0.13`, 1 NVIDIA GPU
- Total: 2 GPUs
### Model Specifications
`Qwen2.5-7B-Instruct` requires approximately 14GB VRAM in FP16. With `--tp 2`, each GPU needs about 7GB (weights) + 2-3GB (KV cache).
### Network Configuration
Nodes communicate via Ethernet (TCP), using the eno1 network interface.
> Note: Check your actual interface using ip addr command
### Precision
Using FP16 precision to maintain maximum accuracy, resulting in higher VRAM usage that requires optimization.
## 2. Prerequisites
Ensure the following requirements are met before installation and deployment:
### Operating System
- Recommended: Ubuntu 20.04/22.04 or other Linux distributions (Windows not recommended, requires WSL2)
- Consistent environments across nodes preferred, though OS can differ if Python environments match
### Network Connectivity
- Node 0 (192.168.0.12) and Node 1 (192.168.0.13) must be able to ping each other:
```shell
ping 192.168.0.12 # from Node 1
ping 192.168.0.13 # from Node 0
```
- Ports 50000 (distributed initialization) and 30000 (HTTP server) must not be blocked by firewall:
```bash
sudo ufw allow 50000
sudo ufw allow 30000
```
- Verify network interface eno1:
```bash
# Adjust interface name as needed
ip addr show eno1
```
If `eno1` doesn't exist, use your actual interface (e.g., `eth0` or `enp0s3`).
### GPU Drivers and CUDA
- Install NVIDIA drivers (version ≥ 470) and `CUDA Toolkit` (12.x recommended):
```bash
nvidia-smi # verify driver and CUDA version
```
Output should show NVIDIA and CUDA versions (e.g., 12.4).
If not installed, refer to NVIDIA's official website for installation.
### Python Environment
- Python 3.9+ (3.10 recommended)
- Consistent Python versions across nodes:
```bash
python3 --version
```
### Disk Space
- `Qwen2.5-7B-Instruct` model requires approximately 15GB disk space
- Ensure sufficient space in `/opt/models/Qwen/Qwen2.5-7B-Instruct` path
## 3. Installing SGLang
Install SGLang and dependencies on both nodes. Execute the following steps on each computer.
### 3.1 Create Virtual Environment (conda)
```bash
conda create -n sglang_env python=3.10
conda activate sglang_env
```
### 3.2 Install SGLang
> Note: Installation will automatically include GPU-related dependencies like `torch`, `transformers`, `flashinfer`
```bash
pip install --upgrade pip
pip install uv
uv pip install "sglang[all]>=0.4.5" --find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer-python
```
Verify installation:
```bash
python -m sglang.launch_server --help
```
Should display SGLang's command-line parameter help information.
### 3.3 Download Qwen2.5-7B-Instruct Model
> Use `huggingface` internationally, `modelscope` within China
Download the model to the same path on both nodes (e.g., `/opt/models/Qwen/Qwen2.5-7B-Instruct`):
```bash
pip install modelscope
modelscope download Qwen/Qwen2.5-7B-Instruct --local-dir /opt/models/Qwen/Qwen2.5-7B-Instruct
```
Alternatively, manually download from `Hugging Face` or `modelscope` and extract to the specified path. Ensure model files are identical across nodes.
## 4. Configuring Dual-Node Deployment
Use tensor parallelism (--tp 2) to distribute the model across 2 GPUs (one per node). Below are the detailed deployment steps and commands.
### 4.1 Deployment Commands
- Node 0 (IP: 192.168.0.12):
```bash
NCCL_IB_DISABLE=1 NCCL_P2P_DISABLE=1 GLOO_SOCKET_IFNAME=eno1 NCCL_SOCKET_IFNAME=eno1 python3 -m sglang.launch_server \
--model-path /opt/models/Qwen/Qwen2.5-7B-Instruct \
--tp 2 \
--nnodes 2 \
--node-rank 0 \
--dist-init-addr 192.168.0.12:50000 \
--disable-cuda-graph \
--host 0.0.0.0 \
--port 30000 \
--mem-fraction-static 0.7
```
- Node 1 (IP: 192.168.0.13):
```bash
NCCL_IB_DISABLE=1 NCCL_P2P_DISABLE=1 GLOO_SOCKET_IFNAME=eno1 NCCL_SOCKET_IFNAME=eno1 python3 -m sglang.launch_server \
--model-path /opt/models/Qwen/Qwen2.5-7B-Instruct \
--tp 2 \
--nnodes 2 \
--node-rank 1 \
--dist-init-addr 192.168.0.12:50000 \
--disable-cuda-graph \
--host 0.0.0.0 \
--port 30000 \
--mem-fraction-static 0.7
```
> Note: If OOM occurs, adjust the `--mem-fraction-static` parameter from the default 0.9 to 0.7. This change reduces VRAM usage by about 2GB for the current 7B model.
> CUDA Graph allocates additional VRAM (typically hundreds of MB) to store computation graphs. If VRAM is near capacity, enabling CUDA Graph may trigger OOM errors.
## Additional Parameters and Information
**[Original Article](https://stable-learn.com/en/sglang-muti-nodes-deploy/)**
| 2025-04-18T14:55:00 |
https://www.reddit.com/r/LocalLLaMA/comments/1k26vvg/multinode_cluster_deployment_of_qwen_series/
|
suke-wangsr
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k26vvg
| false | null |
t3_1k26vvg
|
/r/LocalLLaMA/comments/1k26vvg/multinode_cluster_deployment_of_qwen_series/
| false | false |
self
| 4 |
{'enabled': False, 'images': [{'id': 'd3age5M-eenRm3-Om0rVIaZy55wZsf3H8AAXiu3vAlQ', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/iaRl_OVFS73txV_NNDbFxeKO_1LqH8u8KefHeoel9uM.jpg?width=108&crop=smart&auto=webp&s=6e13a75e2b9ba8f5329742d85c2c5ed93bb29bcb', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/iaRl_OVFS73txV_NNDbFxeKO_1LqH8u8KefHeoel9uM.jpg?width=216&crop=smart&auto=webp&s=35d6575558fcc623dc583297c6706cca51d3fe15', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/iaRl_OVFS73txV_NNDbFxeKO_1LqH8u8KefHeoel9uM.jpg?width=320&crop=smart&auto=webp&s=cd65ebb11110853b0426b752dbbcdc8be236a18e', 'width': 320}, {'height': 480, 'url': 'https://external-preview.redd.it/iaRl_OVFS73txV_NNDbFxeKO_1LqH8u8KefHeoel9uM.jpg?width=640&crop=smart&auto=webp&s=ce263a128e504263eeb00de624513b089421db5f', 'width': 640}, {'height': 720, 'url': 'https://external-preview.redd.it/iaRl_OVFS73txV_NNDbFxeKO_1LqH8u8KefHeoel9uM.jpg?width=960&crop=smart&auto=webp&s=57980b07449bb9fa5422bdd9b7ede4d3da7e61d3', 'width': 960}, {'height': 810, 'url': 'https://external-preview.redd.it/iaRl_OVFS73txV_NNDbFxeKO_1LqH8u8KefHeoel9uM.jpg?width=1080&crop=smart&auto=webp&s=82206e77361c698c706a27d5f8155a186c1fbd2e', 'width': 1080}], 'source': {'height': 1296, 'url': 'https://external-preview.redd.it/iaRl_OVFS73txV_NNDbFxeKO_1LqH8u8KefHeoel9uM.jpg?auto=webp&s=e4e6979220fc3da29407c5bf045ad471782491a5', 'width': 1728}, 'variants': {}}]}
|
LM Studio error for large models
| 1 |
[removed]
| 2025-04-18T15:08:15 |
https://www.reddit.com/r/LocalLLaMA/comments/1k277at/lm_studio_error_for_large_models/
|
Dazzling-Try-7499
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k277at
| false | null |
t3_1k277at
|
/r/LocalLLaMA/comments/1k277at/lm_studio_error_for_large_models/
| false | false |
self
| 1 | null |
I created an interactive tool to visualize *every* attention weight matrix within GPT-2!
| 252 | 2025-04-18T15:18:17 |
https://v.redd.it/dgo9qamv0mve1
|
tycho_brahes_nose_
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k27fz2
| false |
{'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/dgo9qamv0mve1/DASHPlaylist.mpd?a=1747584481%2CN2MzNDE2MDZhYWUyMTBjNDY4MzA3NjE5OGYzN2E2ZTQyNjE5OTFjNjk3Y2QyODk3YjFjYTI3NDgwZGZmY2YyOA%3D%3D&v=1&f=sd', 'duration': 24, 'fallback_url': 'https://v.redd.it/dgo9qamv0mve1/DASH_1080.mp4?source=fallback', 'has_audio': False, 'height': 1080, 'hls_url': 'https://v.redd.it/dgo9qamv0mve1/HLSPlaylist.m3u8?a=1747584481%2CMGZhYmEwNzJjM2ViODQwZjdhZWZjNTI0NDU5MDdmMDI1NThmNzcxM2E2NDE5YThlNjRmYmYwZWFhNmIyOTYyNA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/dgo9qamv0mve1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}}
|
t3_1k27fz2
|
/r/LocalLLaMA/comments/1k27fz2/i_created_an_interactive_tool_to_visualize_every/
| false | false | 252 |
{'enabled': False, 'images': [{'id': 'YW45M2FibXYwbXZlMWaepLM_4Oin4KjR_zAxiUwp5NOaLzCHkxa3urw0ZqL6', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/YW45M2FibXYwbXZlMWaepLM_4Oin4KjR_zAxiUwp5NOaLzCHkxa3urw0ZqL6.png?width=108&crop=smart&format=pjpg&auto=webp&s=199b275a949f0e92d669987a13fc4cc556875033', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/YW45M2FibXYwbXZlMWaepLM_4Oin4KjR_zAxiUwp5NOaLzCHkxa3urw0ZqL6.png?width=216&crop=smart&format=pjpg&auto=webp&s=7543a7f3aa1191bea128907544d635af9aa4fe2c', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/YW45M2FibXYwbXZlMWaepLM_4Oin4KjR_zAxiUwp5NOaLzCHkxa3urw0ZqL6.png?width=320&crop=smart&format=pjpg&auto=webp&s=fa7571ee45adcf9c8c45e13a141f103d447abdaa', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/YW45M2FibXYwbXZlMWaepLM_4Oin4KjR_zAxiUwp5NOaLzCHkxa3urw0ZqL6.png?width=640&crop=smart&format=pjpg&auto=webp&s=1d4f107bd9adcda43534f52889dcb4ff20c46f12', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/YW45M2FibXYwbXZlMWaepLM_4Oin4KjR_zAxiUwp5NOaLzCHkxa3urw0ZqL6.png?width=960&crop=smart&format=pjpg&auto=webp&s=50d511f4a4b4076c4112f8767f9dc27a92f4fecf', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/YW45M2FibXYwbXZlMWaepLM_4Oin4KjR_zAxiUwp5NOaLzCHkxa3urw0ZqL6.png?width=1080&crop=smart&format=pjpg&auto=webp&s=685ccb5b4d9b9fe080ac4b2f2a46408eaecbb446', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/YW45M2FibXYwbXZlMWaepLM_4Oin4KjR_zAxiUwp5NOaLzCHkxa3urw0ZqL6.png?format=pjpg&auto=webp&s=1cb78bfc28d5c817b9e2b7ab5240b1dea75ef93a', 'width': 1920}, 'variants': {}}]}
|
||
What is your cheapest LocalLLAMA portable setup?
| 1 |
[removed]
| 2025-04-18T15:22:25 |
https://www.reddit.com/r/LocalLLaMA/comments/1k27jng/what_is_your_cheapest_localllama_portable_setup/
|
0xfleventy5
|
self.LocalLLaMA
| 2025-04-18T15:31:27 | 0 |
{}
|
1k27jng
| false | null |
t3_1k27jng
|
/r/LocalLLaMA/comments/1k27jng/what_is_your_cheapest_localllama_portable_setup/
| false | false |
self
| 1 | null |
Playing DOOM II and 19 other DOS/GB games with LLMs as a new benchmark
| 1 |
[removed]
| 2025-04-18T15:49:40 |
https://v.redd.it/boqzmm4y6mve1
|
Top_Courage_844
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k286vj
| false |
{'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/boqzmm4y6mve1/DASHPlaylist.mpd?a=1747584906%2COWVlOWI0YzFiMGY0ZDBiOWIyYWFjZjlkODBmNzQxYWY0ZmY0ODQwMGQ1MTFmNGYxNWVkMzkxMTEzZmU0NTM4Yw%3D%3D&v=1&f=sd', 'duration': 92, 'fallback_url': 'https://v.redd.it/boqzmm4y6mve1/DASH_720.mp4?source=fallback', 'has_audio': False, 'height': 720, 'hls_url': 'https://v.redd.it/boqzmm4y6mve1/HLSPlaylist.m3u8?a=1747584906%2CMWFjYjg0NGExMWZhNWQ1ZmVkNDdlMWYyYmUyMjdlZmY5OTliY2VlODNiYzdmZDI0ZjEzODQwZjdjMjQ2OWFkNA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/boqzmm4y6mve1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1062}}
|
t3_1k286vj
|
/r/LocalLLaMA/comments/1k286vj/playing_doom_ii_and_19_other_dosgb_games_with/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'cjNvZnVrNHk2bXZlMY3P1Exc5u-yEn5QEnlfLIhdgOY4HLTGRGU3oRQL_0lq', 'resolutions': [{'height': 73, 'url': 'https://external-preview.redd.it/cjNvZnVrNHk2bXZlMY3P1Exc5u-yEn5QEnlfLIhdgOY4HLTGRGU3oRQL_0lq.png?width=108&crop=smart&format=pjpg&auto=webp&s=992b0e3d1d24a2ac29a7f010d88308223ee0b7c3', 'width': 108}, {'height': 146, 'url': 'https://external-preview.redd.it/cjNvZnVrNHk2bXZlMY3P1Exc5u-yEn5QEnlfLIhdgOY4HLTGRGU3oRQL_0lq.png?width=216&crop=smart&format=pjpg&auto=webp&s=f381a1c2a95ecf95b3799b74108cfaddabe1ebda', 'width': 216}, {'height': 216, 'url': 'https://external-preview.redd.it/cjNvZnVrNHk2bXZlMY3P1Exc5u-yEn5QEnlfLIhdgOY4HLTGRGU3oRQL_0lq.png?width=320&crop=smart&format=pjpg&auto=webp&s=30c189a612344751dfab6b55a2389587e7637bde', 'width': 320}, {'height': 433, 'url': 'https://external-preview.redd.it/cjNvZnVrNHk2bXZlMY3P1Exc5u-yEn5QEnlfLIhdgOY4HLTGRGU3oRQL_0lq.png?width=640&crop=smart&format=pjpg&auto=webp&s=75aa378ebf2b542261ffb3e3af9b3433ac20a16c', 'width': 640}, {'height': 650, 'url': 'https://external-preview.redd.it/cjNvZnVrNHk2bXZlMY3P1Exc5u-yEn5QEnlfLIhdgOY4HLTGRGU3oRQL_0lq.png?width=960&crop=smart&format=pjpg&auto=webp&s=ce54c144a0457ae030e1605ae9404eb80713bcfc', 'width': 960}, {'height': 732, 'url': 'https://external-preview.redd.it/cjNvZnVrNHk2bXZlMY3P1Exc5u-yEn5QEnlfLIhdgOY4HLTGRGU3oRQL_0lq.png?width=1080&crop=smart&format=pjpg&auto=webp&s=dcbca8458681e5f20316d90123f9d2505d8fc309', 'width': 1080}], 'source': {'height': 800, 'url': 'https://external-preview.redd.it/cjNvZnVrNHk2bXZlMY3P1Exc5u-yEn5QEnlfLIhdgOY4HLTGRGU3oRQL_0lq.png?format=pjpg&auto=webp&s=a2204031b79e11429d0be2b504689ca7f825b698', 'width': 1180}, 'variants': {}}]}
|
|
RUN - Compare ChatGPT, DeepSeek, Gemini & 60+ Models For a $1
| 1 |
[removed]
| 2025-04-18T15:56:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1k28cia/run_compare_chatgpt_deepseek_gemini_60_models_for/
|
Buffalo_Emotional
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k28cia
| false | null |
t3_1k28cia
|
/r/LocalLLaMA/comments/1k28cia/run_compare_chatgpt_deepseek_gemini_60_models_for/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'jYq6kalyf0JnrtHalqazQvcmSFvbdfsZaNQ6AqAJCQ8', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/VdPdaXUcD8Q0zmiJPRdkTD9L00ECZeqAdsPMWBdsUlw.jpg?width=108&crop=smart&auto=webp&s=7e4a7b36b4f75a17af8d3d9c90e5d32b820a8aa2', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/VdPdaXUcD8Q0zmiJPRdkTD9L00ECZeqAdsPMWBdsUlw.jpg?width=216&crop=smart&auto=webp&s=ab6f33df681c28fe71b4f4d46ef6469ef28ebfb5', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/VdPdaXUcD8Q0zmiJPRdkTD9L00ECZeqAdsPMWBdsUlw.jpg?width=320&crop=smart&auto=webp&s=b60a4d25e724d6b16e95d3626f117ae8398bfdd9', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/VdPdaXUcD8Q0zmiJPRdkTD9L00ECZeqAdsPMWBdsUlw.jpg?width=640&crop=smart&auto=webp&s=f5159d4791b7abeb471bd9da11f4e4d8a88a6fca', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/VdPdaXUcD8Q0zmiJPRdkTD9L00ECZeqAdsPMWBdsUlw.jpg?width=960&crop=smart&auto=webp&s=0eee6e1d5c15ba21c1fb1a4f7e6b9bca9c89d9d7', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/VdPdaXUcD8Q0zmiJPRdkTD9L00ECZeqAdsPMWBdsUlw.jpg?width=1080&crop=smart&auto=webp&s=228c39c58cb368426360267470acbf81451f084b', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/VdPdaXUcD8Q0zmiJPRdkTD9L00ECZeqAdsPMWBdsUlw.jpg?auto=webp&s=72296686a6191277739b17e3f180eae0ea98d853', 'width': 1200}, 'variants': {}}]}
|
I tried fine-tuning Qwen2.5 to generate git commit messages
| 21 |
Hi I recently tried fine-tuning Qwen2.5-Coder-3B-Instruct to generate better commit messages. The main goal is to let it understand the idea behind code changes instead of simply repeating them. Qwen2.5-Coder-3B-Instruct is a sweet model that is capable in coding tasks and lightweight to run. Then, I fine tune it on the dataset [Maxscha/commitbench](https://huggingface.co/datasets/Maxscha/commitbench).
I think the results are honestly not bad. If the code changes focus on a main goal, the model can guess it pretty well. I released it as a python package and it is available now. You may check the fine tune script to see the training details as well. Hope you find them useful.
You can use it by first installing `pip install git-gen-utils` and running `git-gen`
🔗Source: [https://github.com/CyrusCKF/git-gen](https://github.com/CyrusCKF/git-gen)
🤖Script: [https://github.com/CyrusCKF/git-gen/blob/main/finetune/finetune.ipynb](https://github.com/CyrusCKF/git-gen/blob/main/finetune/finetune.ipynb)
🤗Model (on HuggingFace): [https://huggingface.co/CyrusCheungkf/git-commit-3B](https://huggingface.co/CyrusCheungkf/git-commit-3B)
| 2025-04-18T15:56:26 |
https://www.reddit.com/r/LocalLLaMA/comments/1k28cqz/i_tried_finetuning_qwen25_to_generate_git_commit/
|
m19990328
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k28cqz
| false | null |
t3_1k28cqz
|
/r/LocalLLaMA/comments/1k28cqz/i_tried_finetuning_qwen25_to_generate_git_commit/
| false | false |
self
| 21 |
{'enabled': False, 'images': [{'id': 'xy6I31nAkkoL8Ct3IvkIG0o5o6lAGtSqUep7-mqoHjI', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/Slm4deth7Q2I29iTzhg7J_ZJ8adLRcnJfm_APLi_q2o.jpg?width=108&crop=smart&auto=webp&s=21a01be714d068616d5420f0c3260236e096005b', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/Slm4deth7Q2I29iTzhg7J_ZJ8adLRcnJfm_APLi_q2o.jpg?width=216&crop=smart&auto=webp&s=07f4432d9ac7b143a81621fd7d406b5c6c6ba5a5', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/Slm4deth7Q2I29iTzhg7J_ZJ8adLRcnJfm_APLi_q2o.jpg?width=320&crop=smart&auto=webp&s=122c7bd960bbf3e4d5623e70a53ecf84813a8600', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/Slm4deth7Q2I29iTzhg7J_ZJ8adLRcnJfm_APLi_q2o.jpg?width=640&crop=smart&auto=webp&s=613ff7deda28e7cf2a9f51c5d21978d338f7cdfe', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/Slm4deth7Q2I29iTzhg7J_ZJ8adLRcnJfm_APLi_q2o.jpg?width=960&crop=smart&auto=webp&s=355dc331548f103574b20bdd52bbc96cde725054', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/Slm4deth7Q2I29iTzhg7J_ZJ8adLRcnJfm_APLi_q2o.jpg?width=1080&crop=smart&auto=webp&s=40cca150b01e580fb9bfc0ec6859dc345ed7cfa5', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/Slm4deth7Q2I29iTzhg7J_ZJ8adLRcnJfm_APLi_q2o.jpg?auto=webp&s=1f8e7207784a222a3af9406ded1d95cd89633695', 'width': 1200}, 'variants': {}}]}
|
Playing DOOM II and 19 other DOS/GB games with LLMs as a new benchmark
| 898 |
From AK (@akhaliq)
"We introduce a research preview of VideoGameBench, a benchmark which challenges vision-language models to complete, in real-time, a suite of 20 different popular video games from both hand-held consoles and PC
GPT-4o, Claude Sonnet 3.7, Gemini 2.5 Pro, and Gemini 2.0 Flash playing Doom II (default difficulty) on VideoGameBench-Lite with the same input prompt! Models achieve varying levels of success but none are able to pass even the first level."
project page: [https://vgbench.com](https://vgbench.com)
try on other games: [https://github.com/alexzhang13/VideoGameBench](https://github.com/alexzhang13/VideoGameBench)
| 2025-04-18T15:59:16 |
https://v.redd.it/u1i2op2o8mve1
|
ZhalexDev
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k28f3f
| false |
{'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/u1i2op2o8mve1/DASHPlaylist.mpd?a=1747585072%2CMTc0NmNkMGQ5MTVmZGRjZTMxZmE4ZjEzMzg4M2MzM2EwZDYyYzE2Zjc2YTUzMzlkNTY2NDI0ZWRjZGEyMmE2MA%3D%3D&v=1&f=sd', 'duration': 92, 'fallback_url': 'https://v.redd.it/u1i2op2o8mve1/DASH_720.mp4?source=fallback', 'has_audio': False, 'height': 720, 'hls_url': 'https://v.redd.it/u1i2op2o8mve1/HLSPlaylist.m3u8?a=1747585072%2COTE2MjJmZGNlYjhmMzM4ODlkZmZjODg3MmI5NGIwOTRiYjlkZGQwODRmZmIxM2M3MDY0NTk2OGNmN2MyNjQxMw%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/u1i2op2o8mve1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1062}}
|
t3_1k28f3f
|
/r/LocalLLaMA/comments/1k28f3f/playing_doom_ii_and_19_other_dosgb_games_with/
| false | false | 898 |
{'enabled': False, 'images': [{'id': 'd3J6N2xwMm84bXZlMeIZf5sR-oXFPwhpDTHMtN-Je-w0GMxJeu96UcIYpm6F', 'resolutions': [{'height': 73, 'url': 'https://external-preview.redd.it/d3J6N2xwMm84bXZlMeIZf5sR-oXFPwhpDTHMtN-Je-w0GMxJeu96UcIYpm6F.png?width=108&crop=smart&format=pjpg&auto=webp&s=d4144cefefe046ae689420f48d3bd13856c3022c', 'width': 108}, {'height': 146, 'url': 'https://external-preview.redd.it/d3J6N2xwMm84bXZlMeIZf5sR-oXFPwhpDTHMtN-Je-w0GMxJeu96UcIYpm6F.png?width=216&crop=smart&format=pjpg&auto=webp&s=a884ee00c9797aaac422a1a29440348e45ee55c0', 'width': 216}, {'height': 216, 'url': 'https://external-preview.redd.it/d3J6N2xwMm84bXZlMeIZf5sR-oXFPwhpDTHMtN-Je-w0GMxJeu96UcIYpm6F.png?width=320&crop=smart&format=pjpg&auto=webp&s=3473ca6cc67944bea65dcf559c0c53185d9b00a5', 'width': 320}, {'height': 433, 'url': 'https://external-preview.redd.it/d3J6N2xwMm84bXZlMeIZf5sR-oXFPwhpDTHMtN-Je-w0GMxJeu96UcIYpm6F.png?width=640&crop=smart&format=pjpg&auto=webp&s=bbe06342a2a97732985538d5262e84484ed17cd9', 'width': 640}, {'height': 650, 'url': 'https://external-preview.redd.it/d3J6N2xwMm84bXZlMeIZf5sR-oXFPwhpDTHMtN-Je-w0GMxJeu96UcIYpm6F.png?width=960&crop=smart&format=pjpg&auto=webp&s=400badfb931ad0368c69756a50b0808d1f167eda', 'width': 960}, {'height': 732, 'url': 'https://external-preview.redd.it/d3J6N2xwMm84bXZlMeIZf5sR-oXFPwhpDTHMtN-Je-w0GMxJeu96UcIYpm6F.png?width=1080&crop=smart&format=pjpg&auto=webp&s=65482d4c29c833e13aa98fa4d3efabbd274c3c8d', 'width': 1080}], 'source': {'height': 800, 'url': 'https://external-preview.redd.it/d3J6N2xwMm84bXZlMeIZf5sR-oXFPwhpDTHMtN-Je-w0GMxJeu96UcIYpm6F.png?format=pjpg&auto=webp&s=650cf8210f792538745698e02b8afba1695e867b', 'width': 1180}, 'variants': {}}]}
|
|
I wish there were a latex and/or tikz benchmark
| 1 |
I don't even have a GPU so there is no chance I could run the tests. But if anyone is interested, I could maybe help ask the questions.
| 2025-04-18T15:59:55 |
https://www.reddit.com/r/LocalLLaMA/comments/1k28fnl/i_wish_there_were_a_latex_andor_tikz_benchmark/
|
MrMrsPotts
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k28fnl
| false | null |
t3_1k28fnl
|
/r/LocalLLaMA/comments/1k28fnl/i_wish_there_were_a_latex_andor_tikz_benchmark/
| false | false |
self
| 1 | null |
Llama 4 Maverick MLX performance on M3 Ultra
| 29 |
LM studio released an MLX update today so we can run Maverick in MLX format.
Q4 version numbers:
Prompt size: 12405
Prompt eval rate: 332 t/s
Token gen rate: 47.42
Right now for me there is a bug where it's not using prompt caching. Promising initial results though.
| 2025-04-18T16:03:30 |
https://www.reddit.com/r/LocalLLaMA/comments/1k28j02/llama_4_maverick_mlx_performance_on_m3_ultra/
|
nomorebuttsplz
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k28j02
| false | null |
t3_1k28j02
|
/r/LocalLLaMA/comments/1k28j02/llama_4_maverick_mlx_performance_on_m3_ultra/
| false | false |
self
| 29 | null |
Time to step up the /local reasoning game
| 327 |
Latest OAI models tucked away behind intrusive "ID verification"....
| 2025-04-18T16:17:11 |
vornamemitd
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1k28ulo
| false | null |
t3_1k28ulo
|
/r/LocalLLaMA/comments/1k28ulo/time_to_step_up_the_local_reasoning_game/
| false | false | 327 |
{'enabled': True, 'images': [{'id': 'G2W4Bxj2_GLR0tp5DlTpjyCug_FvDl74xFGlre16JEA', 'resolutions': [{'height': 64, 'url': 'https://preview.redd.it/wtibm8c3cmve1.jpeg?width=108&crop=smart&auto=webp&s=5feb2724acfab316ff5a490a6031f61c2b3ee110', 'width': 108}, {'height': 129, 'url': 'https://preview.redd.it/wtibm8c3cmve1.jpeg?width=216&crop=smart&auto=webp&s=153f4d64e647ddd0589d29ce571a795d8dc22b9d', 'width': 216}, {'height': 191, 'url': 'https://preview.redd.it/wtibm8c3cmve1.jpeg?width=320&crop=smart&auto=webp&s=e5d0bdad113d36d5c64c0bdb2df8b1e4215646dd', 'width': 320}, {'height': 382, 'url': 'https://preview.redd.it/wtibm8c3cmve1.jpeg?width=640&crop=smart&auto=webp&s=93f80a0bad3e3f79619d29663e49d519eaa7898d', 'width': 640}, {'height': 574, 'url': 'https://preview.redd.it/wtibm8c3cmve1.jpeg?width=960&crop=smart&auto=webp&s=5d07a1b7958059279032a275ed0d7168d9da0faf', 'width': 960}, {'height': 646, 'url': 'https://preview.redd.it/wtibm8c3cmve1.jpeg?width=1080&crop=smart&auto=webp&s=b34018c078a4e1b303a6338fe64c6b947fb25090', 'width': 1080}], 'source': {'height': 718, 'url': 'https://preview.redd.it/wtibm8c3cmve1.jpeg?auto=webp&s=be9362f02c4338ab26a4f83156119c3de345395a', 'width': 1200}, 'variants': {}}]}
|
||
Docker desktop now supports model running
| 0 |
Didn't see a post here yet... Anyone try it yet? Thoughts? https://www.docker.com/blog/introducing-docker-model-runner/
| 2025-04-18T16:21:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1k28y2j/docker_desktop_now_supports_model_running/
|
onemoreburrito
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k28y2j
| false | null |
t3_1k28y2j
|
/r/LocalLLaMA/comments/1k28y2j/docker_desktop_now_supports_model_running/
| false | false |
self
| 0 |
{'enabled': False, 'images': [{'id': '5YurRdfkQeIxTtv_1yZqhwCibgtikSBDEaMPw2UzacA', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/Gz5iv0hC1-oojGEzY0yl1Njb0Jt7-bSeQ06GaO9A-dI.jpg?width=108&crop=smart&auto=webp&s=0f8b485a55e05dff7858656d6ba29a819f1a1fb1', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/Gz5iv0hC1-oojGEzY0yl1Njb0Jt7-bSeQ06GaO9A-dI.jpg?width=216&crop=smart&auto=webp&s=3af5ef04002ac63ee54e91eecec3e7530b47e145', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/Gz5iv0hC1-oojGEzY0yl1Njb0Jt7-bSeQ06GaO9A-dI.jpg?width=320&crop=smart&auto=webp&s=da4a6dc73322d752eb9ee5abefddc500ca1181a8', 'width': 320}, {'height': 640, 'url': 'https://external-preview.redd.it/Gz5iv0hC1-oojGEzY0yl1Njb0Jt7-bSeQ06GaO9A-dI.jpg?width=640&crop=smart&auto=webp&s=001f2b11c0e4f2b2b8e0650342013fc0349a649a', 'width': 640}, {'height': 960, 'url': 'https://external-preview.redd.it/Gz5iv0hC1-oojGEzY0yl1Njb0Jt7-bSeQ06GaO9A-dI.jpg?width=960&crop=smart&auto=webp&s=605af0ce456aecc217b7f2c42c158e04f56cce23', 'width': 960}], 'source': {'height': 1024, 'url': 'https://external-preview.redd.it/Gz5iv0hC1-oojGEzY0yl1Njb0Jt7-bSeQ06GaO9A-dI.jpg?auto=webp&s=93145b74d2c28e71d64929f9373613f99f937937', 'width': 1024}, 'variants': {}}]}
|
Title: Evolution of a Build – ThoughtPenAI’s Super Intelligence Pathway
| 1 |
[removed]
| 2025-04-18T16:21:49 |
https://www.reddit.com/r/LocalLLaMA/comments/1k28yh0/title_evolution_of_a_build_thoughtpenais_super/
|
ArchitectExecutor
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k28yh0
| false | null |
t3_1k28yh0
|
/r/LocalLLaMA/comments/1k28yh0/title_evolution_of_a_build_thoughtpenais_super/
| false | false |
self
| 1 | null |
Simple tikz test . chatgpt o1 passes but grok/gemini 2.5 fail
| 0 |
Prompt: Give me tikz code to draw a heptagon with 20 balls in it. The balls should have different colors and the numbers 1 to 20 written inside them. You should give a full LaTeX document.
chatgpt o1's answer renders as:
https://preview.redd.it/w2u38zkeemve1.png?width=270&format=png&auto=webp&s=d6d784c86455a17d415bc3a37be0fcd471c4608b
What does your favorite local model give?
| 2025-04-18T16:30:10 |
https://www.reddit.com/r/LocalLLaMA/comments/1k295ir/simple_tikz_test_chatgpt_o1_passes_but_grokgemini/
|
MrMrsPotts
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k295ir
| false | null |
t3_1k295ir
|
/r/LocalLLaMA/comments/1k295ir/simple_tikz_test_chatgpt_o1_passes_but_grokgemini/
| false | false | 0 | null |
|
How to add tokens/sec in Enchanted LLM?
| 1 |
[removed]
| 2025-04-18T16:34:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1k299ad/how_to_add_tokenssec_in_enchanted_llm/
|
mykeschn
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k299ad
| false | null |
t3_1k299ad
|
/r/LocalLLaMA/comments/1k299ad/how_to_add_tokenssec_in_enchanted_llm/
| false | false |
self
| 1 | null |
QAT is slowly becoming mainstream now?
| 192 |
Google just released a QAT optimized Gemma 3 - 27 billion parameter model. The quantization aware training claims to recover close to 97% of the accuracy loss that happens during the quantization. Do you think this is slowly becoming the norm? Will non-quantized safetensors slowly become obsolete?
| 2025-04-18T16:52:07 |
https://www.reddit.com/r/LocalLLaMA/comments/1k29oe2/qat_is_slowly_becoming_mainstream_now/
|
__amberluz__
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k29oe2
| false | null |
t3_1k29oe2
|
/r/LocalLLaMA/comments/1k29oe2/qat_is_slowly_becoming_mainstream_now/
| false | false |
self
| 192 | null |
Best Local LLM per 3090 card
| 1 |
[removed]
| 2025-04-18T17:09:29 |
https://www.reddit.com/r/LocalLLaMA/comments/1k2a3tz/best_local_llm_per_3090_card/
|
PTCAH
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1k2a3tz
| false | null |
t3_1k2a3tz
|
/r/LocalLLaMA/comments/1k2a3tz/best_local_llm_per_3090_card/
| false | false |
self
| 1 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.