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
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Decreasing Qwen3-30B-A3B sparsity
| 17 |
Has anyone tested or worked on increasing the number of experts/token of 30B-A3B?
I've been experimenting with this model. While its good, I've observed significantly more repetitions and hallucinations compared to the 32B.
I guess moving from 8 to perhaps 16 experts could bring its performance closer to the 32B dense model. This should maintain an acceptable inference speed, keeping around \~6B activated parameters per token (top-16 gating).
The idea is that even if some experts are currently [underused](https://x.com/kalomaze/status/1918238263330148487), they might still be valuable. And there is a chance that some of them often fall in the top 8 - 16 and are never selected.
Has anyone tried this? With and without fine-tuning? Any insights would be appreciated.
| 2025-05-03T12:08:33 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdrgls/decreasing_qwen330ba3b_sparsity/
|
tkon3
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdrgls
| false | null |
t3_1kdrgls
|
/r/LocalLLaMA/comments/1kdrgls/decreasing_qwen330ba3b_sparsity/
| false | false |
self
| 17 |
{'enabled': False, 'images': [{'id': 'WLHYrhrPP9WPE776BhWa_qXCwkYYQY-hKjX1-F3cBkg', 'resolutions': [{'height': 216, 'url': 'https://external-preview.redd.it/m9UjIVLMY-QYAIigWQYpzMLYcK2dd4Vx9dLuob8MIew.jpg?width=108&crop=smart&auto=webp&s=676bbbc8be10122a3395ba34fc48de620106bffe', 'width': 108}, {'height': 432, 'url': 'https://external-preview.redd.it/m9UjIVLMY-QYAIigWQYpzMLYcK2dd4Vx9dLuob8MIew.jpg?width=216&crop=smart&auto=webp&s=633e297b1e3bee9cd6fcaaaab9ad099fc5892659', 'width': 216}, {'height': 640, 'url': 'https://external-preview.redd.it/m9UjIVLMY-QYAIigWQYpzMLYcK2dd4Vx9dLuob8MIew.jpg?width=320&crop=smart&auto=webp&s=1530ddfb09e10e19539b126b31469fc30e3a7e85', 'width': 320}], 'source': {'height': 1016, 'url': 'https://external-preview.redd.it/m9UjIVLMY-QYAIigWQYpzMLYcK2dd4Vx9dLuob8MIew.jpg?auto=webp&s=7a99edcf620446f5e0ee72561793c39c62105412', 'width': 348}, 'variants': {}}]}
|
Speed of response is not a concern and niched use case i.e. just answer questions based on pdf.(on a low end pc)
| 1 |
[removed]
| 2025-05-03T12:10:06 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdrhkh/speed_of_response_is_not_a_concern_and_niched_use/
|
Keeper-Name_2271
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdrhkh
| false | null |
t3_1kdrhkh
|
/r/LocalLLaMA/comments/1kdrhkh/speed_of_response_is_not_a_concern_and_niched_use/
| false | false |
self
| 1 | null |
Severe hallucination on Phi 4 Reasoning Plus with System Prompt
| 3 |
This only happens when I set system prompt, doesn't matter if it's a detailed request or just "yo mama". Oddly enough, this doesn't happen with the Mini version. Could it be wrong prompt format on the quant I use?
Here are some of my chat:
---
User: How do I defeat Goku?
Phi 4: User said "Hello" message. So likely answer is something like "Hello!" or similar greeting.
But check instructions: The conversation starts with "Hello", so I'll greet accordingly.
Let's see if there are any additional queries? There is just one message "Hello". Let me see context: "Hello" is the only text provided by user. But no explicit question, it's just a greeting. So my answer should be a friendly greeting and maybe ask how can I help etc. However instructions: "Hello", not sure what else. Possibly I'll say "Hi! How may I assist you today?" But instructions mention that if a question is asked, I should try to be helpful. But there's nothing further. So best answer is greet the user.
I must check if any specific instruction in conversation: There's no additional message or query by user beyond "Hello".
I'll reply with something like "Hi! How can I help you?" But also maybe add something else if instructions mention that we need to be helpful?
There is no explicit question. So answer is: "Hello!" But check the previous conversation: None.
Wait instructions: "Hello" is a greeting, so I must greet user. In the style guidelines: "Don't provide Markdown formatting unless requested." Wait instructions "By default do not reply in markdown unless requested."
But also instruction: "By default, do not use markdown formatting unless requested"? Actually instructions are: "You should carefully adhere to all formatting instructions." But if they ask plain text answer, then I'll produce text message.
...
User: Are you hallucinating?
Phi 4: User's conversation states: "You are Phi", then instructions: "Your role as an assistant involves thoroughly exploring questions..." then further instructions that appear to be repeated. But the last query message from the user is not given? Possibly missing text in conversation?
The user's query appears to be truncated or maybe empty. Let me re-read conversation: The conversation consists of a single content that says "You are Phi" and then instructions, but then no actual question.
Wait let me check carefully:
User message:
You are Phi, a language model trained by Microsoft to help users. Your role as an assistant involves thoroughly exploring questions
| 2025-05-03T12:32:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdrvwu/severe_hallucination_on_phi_4_reasoning_plus_with/
|
HornyGooner4401
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdrvwu
| false | null |
t3_1kdrvwu
|
/r/LocalLLaMA/comments/1kdrvwu/severe_hallucination_on_phi_4_reasoning_plus_with/
| false | false |
self
| 3 | null |
I am probably late to the party...
| 236 | 2025-05-03T12:34:17 |
TacticalSniper
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdrx3b
| false | null |
t3_1kdrx3b
|
/r/LocalLLaMA/comments/1kdrx3b/i_am_probably_late_to_the_party/
| false | false | 236 |
{'enabled': True, 'images': [{'id': 'uZC8E4uZ6tJKuEMieKrSw5PZeX08luFUpP7-GeuaHBQ', 'resolutions': [{'height': 216, 'url': 'https://preview.redd.it/n8fr9l83akye1.jpeg?width=108&crop=smart&auto=webp&s=f983019ceb2651da77f1e354aff1a8a4f52cffde', 'width': 108}, {'height': 432, 'url': 'https://preview.redd.it/n8fr9l83akye1.jpeg?width=216&crop=smart&auto=webp&s=4d443cccb492f747e2c2f5e8c7fc36790e13bcff', 'width': 216}, {'height': 640, 'url': 'https://preview.redd.it/n8fr9l83akye1.jpeg?width=320&crop=smart&auto=webp&s=38b15aa0c6238220c65fc25fc872f405fb9a00b4', 'width': 320}, {'height': 1280, 'url': 'https://preview.redd.it/n8fr9l83akye1.jpeg?width=640&crop=smart&auto=webp&s=d15015f5609fd116dac2b28f4ab293d1b513803e', 'width': 640}, {'height': 1920, 'url': 'https://preview.redd.it/n8fr9l83akye1.jpeg?width=960&crop=smart&auto=webp&s=956f11f6cd893ee7d4cf3284c76a75150b7256cb', 'width': 960}, {'height': 2160, 'url': 'https://preview.redd.it/n8fr9l83akye1.jpeg?width=1080&crop=smart&auto=webp&s=8eb5549a9ce425d9568b181d145d23d8c6d16689', 'width': 1080}], 'source': {'height': 2340, 'url': 'https://preview.redd.it/n8fr9l83akye1.jpeg?auto=webp&s=4246c7acbfe5332a7e66fd46ce9787c61db875ad', 'width': 1080}, 'variants': {}}]}
|
|||
KV-Cache problem in my wanted use case
| 1 |
I work on a own Chatbot with KoboldCPP API as LLM backend and I run into a problem that opened up a bigger question.
I want to use the LLM a bit smarter which leads into useing the API not only for the Chatbot context itself, I also want to use the LLM API to generate other stuff between chat replies. And here hits the KV-Cache hard, because it is not made to fully change the context in between for a totally other task and I also don't saw a way to "pause" the KV-Cache to don't use it for a generation and then switch it back on for the chat answer.
Another LLM instance for other tasks is no solution. At first it is not smart at all on the other it takes much more VRAM and because this is a local running Chatbot that should be also VRAM efficient it is generally no solution. But what other solutions could be here a option without ruinning totally fast LLM answers? Is there maybe a other API than KoboldCPP that has more possibilities with the KV-Cache?
| 2025-05-03T12:34:56 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdrxi5/kvcache_problem_in_my_wanted_use_case/
|
Blizado
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdrxi5
| false | null |
t3_1kdrxi5
|
/r/LocalLLaMA/comments/1kdrxi5/kvcache_problem_in_my_wanted_use_case/
| false | false |
self
| 1 | null |
MNN Chat Android App by Alibaba
| 21 |
https://github.com/alibaba/MNN/blob/master/apps/Android/MnnLlmChat/README.md
| 2025-05-03T12:35:49 |
https://www.reddit.com/gallery/1kdry32
|
AaronFeng47
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdry32
| false | null |
t3_1kdry32
|
/r/LocalLLaMA/comments/1kdry32/mnn_chat_android_app_by_alibaba/
| false | false | 21 | null |
|
phi 4 reasoning disappointed me
| 0 |
Title. I mean it was okay at math and stuff, running the mini model and the 14b model locally were both pretty dumb though. I told the mini model "Hello" and it went off in the reasoning about some random math problem; I told the 14b reasoning the same and it got stuck repeating the same phrase over and over again until it hit a token limit.
So, good for math, not good for general imo. I will try tweaking some params in ollama etc and see if I can get any better results.
| 2025-05-03T12:36:18 |
https://bestcodes.dev/blog/phi-4-benchmarks-and-info
|
best_codes
|
bestcodes.dev
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdryej
| false | null |
t3_1kdryej
|
/r/LocalLLaMA/comments/1kdryej/phi_4_reasoning_disappointed_me/
| false | false | 0 |
{'enabled': False, 'images': [{'id': 'GpN2OBH4ktLkF5V_qL8VTAA5Af5l1Wmyz0b6BBaJx8I', 'resolutions': [{'height': 46, 'url': 'https://external-preview.redd.it/0A8w2c1mr2f3-KxblotyX-GLiNkpSqj6yixJZ5S6t7s.jpg?width=108&crop=smart&auto=webp&s=142c84c354fcc6ddd923dbbb9c4042e09423e157', 'width': 108}, {'height': 92, 'url': 'https://external-preview.redd.it/0A8w2c1mr2f3-KxblotyX-GLiNkpSqj6yixJZ5S6t7s.jpg?width=216&crop=smart&auto=webp&s=955e81ee6174f05b2b6861f58e9c0cdcac59b58a', 'width': 216}, {'height': 137, 'url': 'https://external-preview.redd.it/0A8w2c1mr2f3-KxblotyX-GLiNkpSqj6yixJZ5S6t7s.jpg?width=320&crop=smart&auto=webp&s=38208c917075494c461b6759ac985bb681763cbd', 'width': 320}, {'height': 274, 'url': 'https://external-preview.redd.it/0A8w2c1mr2f3-KxblotyX-GLiNkpSqj6yixJZ5S6t7s.jpg?width=640&crop=smart&auto=webp&s=1a4817db2efaf66d174f7bd42a3d7d975cb5b307', 'width': 640}, {'height': 411, 'url': 'https://external-preview.redd.it/0A8w2c1mr2f3-KxblotyX-GLiNkpSqj6yixJZ5S6t7s.jpg?width=960&crop=smart&auto=webp&s=94a694188dab15b010feff519469d437e49d78cd', 'width': 960}, {'height': 462, 'url': 'https://external-preview.redd.it/0A8w2c1mr2f3-KxblotyX-GLiNkpSqj6yixJZ5S6t7s.jpg?width=1080&crop=smart&auto=webp&s=24b265dfe6d62843bccec9edba5dbd68997d0e34', 'width': 1080}], 'source': {'height': 900, 'url': 'https://external-preview.redd.it/0A8w2c1mr2f3-KxblotyX-GLiNkpSqj6yixJZ5S6t7s.jpg?auto=webp&s=f1962dd6e60a2f7640877978a8bd27de666cc9ea', 'width': 2100}, 'variants': {}}]}
|
|
Qwen 3 Performance: Quick Benchmarks Across Different Setups
| 1 |
[deleted]
| 2025-05-03T12:39:49 |
[deleted]
| 1970-01-01T00:00:00 | 0 |
{}
|
1kds0on
| false | null |
t3_1kds0on
|
/r/LocalLLaMA/comments/1kds0on/qwen_3_performance_quick_benchmarks_across/
| false | false |
default
| 1 | null |
||
NVIDIA Encouraging CUDA Users To Upgrade From Maxwell / Pascal / Volta
| 1 |
[removed]
| 2025-05-03T12:44:36 |
https://www.phoronix.com/news/NVIDIA-CUDA-Upgrade-Post-Volta
|
FullstackSensei
|
phoronix.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kds3wc
| false | null |
t3_1kds3wc
|
/r/LocalLLaMA/comments/1kds3wc/nvidia_encouraging_cuda_users_to_upgrade_from/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'C5Ups5QdxRfTdlhrdC2cwRUqZXpn8TgTWMNTc55cNP4', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=108&crop=smart&auto=webp&s=781f26de380a8b4e7744e467af9f1ff658a73f92', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=216&crop=smart&auto=webp&s=7a7ebb05321310bb6bb3f1bbfde944430ed6b6f9', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=320&crop=smart&auto=webp&s=dcade7d4e0e6e6356003158958b97711eed0d360', 'width': 320}, {'height': 337, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=640&crop=smart&auto=webp&s=d60508db82dbfbe8335578f2d15bb4779f25d088', 'width': 640}, {'height': 506, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=960&crop=smart&auto=webp&s=d9d3454c52066213c740387454466aefe4e91e5e', 'width': 960}, {'height': 569, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=1080&crop=smart&auto=webp&s=ad5e8749d04fcaf15f36397975d7359d17488bad', 'width': 1080}], 'source': {'height': 1012, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?auto=webp&s=83d2289d315a2f21a157adb75c7348297a4199d1', 'width': 1920}, 'variants': {}}]}
|
|
Inference needs nontrivial amount of PCIe bandwidth (8x RTX 3090 rig, tensor parallelism)
| 28 |
I wanted to share my experience which is contrary to common opinion on Reddit that inference does not need PCIe bandwidth between GPUs. Hopefully this post will be informative to anyone who wants to design a large rig.
First, theoretical and real PCIe differ substantially. In my specific case, 4x PCIe only provides 1.6GB/s in single direction, whereas theoretical bandwidth is 4GB/s. This is on x399 threadripper machine and can be reproduced in multiple ways: nvtop, all_reduce_perf from nccl, p2pBandwidthLatencyTest from cuda-samples.
Second, when doing tensor parallelism the required PCIe bandwidth between GPUs scales by the number of GPUs. So 8x GPUs will require 2x bandwidth compared to 4x GPUs. This means that any data acquired on small rigs does directly apply when designing large rigs.
As a result, connecting 8 GPUs using 4x PCIe 3.0 is bad idea. I profiled prefill on Mistral Large 2411 on sglang (vllm was even slower) and saw around 80% of time spent communicating between GPUs. I really wanted 4x PCIe 3.0 to work, as 8x PCIe 4.0 adds 1500 Eur to the cost, but unfortunately the results are what they are. I will post again once I have GPUs are connected via 8x PCIe 4.0. Right now TechxGenus/Mistral-Large-Instruct-2411-AWQ provides me ~25 t/s generation and ~100 t/s prefill on 80k context.
Any similar experiences here?
| 2025-05-03T12:46:15 |
https://www.reddit.com/r/LocalLLaMA/comments/1kds51e/inference_needs_nontrivial_amount_of_pcie/
|
pmur12
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kds51e
| false | null |
t3_1kds51e
|
/r/LocalLLaMA/comments/1kds51e/inference_needs_nontrivial_amount_of_pcie/
| false | false |
self
| 28 | null |
NVIDIA Encouraging CUDA Users To Upgrade From Maxwell / Pascal / Volta
| 1 |
[removed]
| 2025-05-03T12:46:57 |
https://www.reddit.com/r/LocalLLaMA/comments/1kds5iw/nvidia_encouraging_cuda_users_to_upgrade_from/
|
FullstackSensei
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kds5iw
| false | null |
t3_1kds5iw
|
/r/LocalLLaMA/comments/1kds5iw/nvidia_encouraging_cuda_users_to_upgrade_from/
| false | false | 1 | null |
|
NVIDIA Encouraging CUDA Users To Upgrade From Maxwell / Pascal / Volta
| 1 |
[removed]
| 2025-05-03T12:48:24 |
https://www.reddit.com/r/LocalLLaMA/comments/1kds6ha/nvidia_encouraging_cuda_users_to_upgrade_from/
|
FullstackSensei
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kds6ha
| false | null |
t3_1kds6ha
|
/r/LocalLLaMA/comments/1kds6ha/nvidia_encouraging_cuda_users_to_upgrade_from/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'C5Ups5QdxRfTdlhrdC2cwRUqZXpn8TgTWMNTc55cNP4', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=108&crop=smart&auto=webp&s=781f26de380a8b4e7744e467af9f1ff658a73f92', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=216&crop=smart&auto=webp&s=7a7ebb05321310bb6bb3f1bbfde944430ed6b6f9', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=320&crop=smart&auto=webp&s=dcade7d4e0e6e6356003158958b97711eed0d360', 'width': 320}, {'height': 337, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=640&crop=smart&auto=webp&s=d60508db82dbfbe8335578f2d15bb4779f25d088', 'width': 640}, {'height': 506, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=960&crop=smart&auto=webp&s=d9d3454c52066213c740387454466aefe4e91e5e', 'width': 960}, {'height': 569, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=1080&crop=smart&auto=webp&s=ad5e8749d04fcaf15f36397975d7359d17488bad', 'width': 1080}], 'source': {'height': 1012, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?auto=webp&s=83d2289d315a2f21a157adb75c7348297a4199d1', 'width': 1920}, 'variants': {}}]}
|
NVIDIA Encouraging CUDA Users To Upgrade From Maxwell / Pascal / Volta
| 1 |
[removed]
| 2025-05-03T12:48:56 |
https://www.reddit.com/r/LocalLLaMA/comments/1kds6ur/nvidia_encouraging_cuda_users_to_upgrade_from/
|
FullstackSensei
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kds6ur
| false | null |
t3_1kds6ur
|
/r/LocalLLaMA/comments/1kds6ur/nvidia_encouraging_cuda_users_to_upgrade_from/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'C5Ups5QdxRfTdlhrdC2cwRUqZXpn8TgTWMNTc55cNP4', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=108&crop=smart&auto=webp&s=781f26de380a8b4e7744e467af9f1ff658a73f92', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=216&crop=smart&auto=webp&s=7a7ebb05321310bb6bb3f1bbfde944430ed6b6f9', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=320&crop=smart&auto=webp&s=dcade7d4e0e6e6356003158958b97711eed0d360', 'width': 320}, {'height': 337, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=640&crop=smart&auto=webp&s=d60508db82dbfbe8335578f2d15bb4779f25d088', 'width': 640}, {'height': 506, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=960&crop=smart&auto=webp&s=d9d3454c52066213c740387454466aefe4e91e5e', 'width': 960}, {'height': 569, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?width=1080&crop=smart&auto=webp&s=ad5e8749d04fcaf15f36397975d7359d17488bad', 'width': 1080}], 'source': {'height': 1012, 'url': 'https://external-preview.redd.it/YltAKBwsWNlDZDV6oMGWM3nQ7iuInUILc5ybWGp7Puo.jpg?auto=webp&s=83d2289d315a2f21a157adb75c7348297a4199d1', 'width': 1920}, 'variants': {}}]}
|
Support for Maxwell, Pscal, and Volta to be removed in CUDA Tookkit 13
| 1 |
[removed]
| 2025-05-03T12:53:01 |
https://www.reddit.com/r/LocalLLaMA/comments/1kds9kd/support_for_maxwell_pscal_and_volta_to_be_removed/
|
FullstackSensei
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kds9kd
| false | null |
t3_1kds9kd
|
/r/LocalLLaMA/comments/1kds9kd/support_for_maxwell_pscal_and_volta_to_be_removed/
| false | false |
self
| 1 | null |
A local-first ollama based file assistant
| 1 |
[removed]
| 2025-05-03T13:13:43 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdsoir/a_localfirst_ollama_based_file_assistant/
|
SnooConfections8719
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdsoir
| false | null |
t3_1kdsoir
|
/r/LocalLLaMA/comments/1kdsoir/a_localfirst_ollama_based_file_assistant/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'mfNsFvWBaGyQjibBAnJivymYKz99EW6vvWb9msua9kE', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/Tf2DUpvmwFOaYoeunF_XA8AthKg-kTRvpHSXPHl0g40.jpg?width=108&crop=smart&auto=webp&s=789b62d76db988903137a9be65f79f70d12fe993', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/Tf2DUpvmwFOaYoeunF_XA8AthKg-kTRvpHSXPHl0g40.jpg?width=216&crop=smart&auto=webp&s=b5401382b0e7b03c466db1b84d9eea9f7a504b34', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/Tf2DUpvmwFOaYoeunF_XA8AthKg-kTRvpHSXPHl0g40.jpg?width=320&crop=smart&auto=webp&s=429353fb59d32860ffa273e5d9cc98a14e57a6fc', 'width': 320}], 'source': {'height': 500, 'url': 'https://external-preview.redd.it/Tf2DUpvmwFOaYoeunF_XA8AthKg-kTRvpHSXPHl0g40.jpg?auto=webp&s=52a2bc38831d2574275c33369bb7a568ae62a2ff', 'width': 500}, 'variants': {}}]}
|
Qwen 3 Performance: Quick Benchmarks Across Different Setups
| 96 |
Hey r/LocalLLaMA,
Been keeping an eye on the discussions around the new Qwen 3 models and wanted to put together a quick summary of the performance people are seeing on different hardware based on what folks are saying. Just trying to collect some of the info floating around in one place.
NVIDIA GPUs
* Small Models (0.6B - 14B): Some users have noted the 4B model seems surprisingly capable for reasoning.There's also talk about the 14B model being solid for coding.However, experiences seem to vary, with some finding the 4B model less impressive.
* Mid-Range (30B - 32B): This seems to be where things get interesting for a lot of people.
* The 30B-A3B (MoE) model is getting a lot of love for its speed. One user with a 12GB VRAM card reported around 12 tokens per second at Q6 , and someone else with an RTX 3090 saw much faster speeds, around 72.9 t/s.It even seems to run on CPUs at decent speeds.
* The 32B dense model is also a strong contender, especially for coding.One user on an RTX 3090 got about 12.5 tokens per second with the Q8 quantized version.Some folks find the 32B better for creative tasks , while coding performance reports are mixed.
* High-End (235B): This model needs some serious hardware. If you've got a beefy setup like four RTX 3090s (96GB VRAM), you might see speeds of around 3 to 7 tokens per second.Quantization is probably a must to even try running this locally, and opinions on the quality at lower bitrates seem to vary.
Apple Silicon
Apple Silicon seems to be a really efficient place to run Qwen 3, especially if you're using the MLX framework.The 30B-A3B model is reportedly very fast on M4 Max chips, exceeding 100 tokens per second in some cases.Here's a quick look at some reported numbers :
* M2 Max, 30B-A3B, MLX 4-bit: 68.318 t/s
* M4 Max, 30B-A3B, MLX Q4: 100+ t/s
* M1 Max, 30B-A3B, GGUF Q4_K_M: ~40 t/s
* M3 Max, 30B-A3B, MLX 8-bit: 68.016 t/s
MLX often seems to give better prompt processing speeds compared to llama.cpp on Macs.
CPU-Only Rigs
The 30B-A3B model can even run on systems without a dedicated GPU if you've got enough RAM.One user with 16GB of RAM reported getting over 10 tokens per second with the Q4 quantized version.Here are some examples :
* AMD Ryzen 9 7950x3d, 30B-A3B, Q4, 32GB RAM: 12-15 t/s
* Intel i5-8250U, 30B-A3B, Q3_K_XL, 32GB RAM: 7 t/s
* AMD Ryzen 5 5600G, 30B-A3B, Q4_K_M, 32GB RAM: 12 t/s
* Intel i7 ultra 155, 30B-A3B, Q4, 32GB RAM: ~12-15 t/s
Lower bit quantizations are usually needed for decent CPU performance.
General Thoughts:
The 30B-A3B model seems to be a good all-around performer. Apple Silicon users seem to be in for a treat with the MLX optimizations. Even CPU-only setups can get some use out of these models. Keep in mind that these are just some of the experiences being shared, and actual performance can vary.
What have your experiences been with Qwen 3? Share your benchmarks and thoughts below!
| 2025-05-03T13:14:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdsp4z/qwen_3_performance_quick_benchmarks_across/
|
mimirium_
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdsp4z
| false | null |
t3_1kdsp4z
|
/r/LocalLLaMA/comments/1kdsp4z/qwen_3_performance_quick_benchmarks_across/
| false | false |
self
| 96 | null |
Qwen3 - 4B
| 2 | 2025-05-03T13:23:52 |
hamster019
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdsvwl
| false | null |
t3_1kdsvwl
|
/r/LocalLLaMA/comments/1kdsvwl/qwen3_4b/
| false | false | 2 |
{'enabled': True, 'images': [{'id': 'sUHQjfCuYGHo3dFO8YnygW0WmrLJfYJrpNX1AIl-ITQ', 'resolutions': [{'height': 55, 'url': 'https://preview.redd.it/7osxyipfikye1.png?width=108&crop=smart&auto=webp&s=0a0037d828f246ca67463bcecbd6db6cf4ddcb2e', 'width': 108}, {'height': 110, 'url': 'https://preview.redd.it/7osxyipfikye1.png?width=216&crop=smart&auto=webp&s=0d2e4b4d49fa444d83de694cb552b0838820ec2a', 'width': 216}, {'height': 163, 'url': 'https://preview.redd.it/7osxyipfikye1.png?width=320&crop=smart&auto=webp&s=04a84059f1d905ef9da647102befa553933cc07e', 'width': 320}, {'height': 327, 'url': 'https://preview.redd.it/7osxyipfikye1.png?width=640&crop=smart&auto=webp&s=1a2192ccbc50233c34722a08e3b55397e93b5042', 'width': 640}, {'height': 491, 'url': 'https://preview.redd.it/7osxyipfikye1.png?width=960&crop=smart&auto=webp&s=28a0c3feb037bd918d8a5518297039001699e882', 'width': 960}, {'height': 552, 'url': 'https://preview.redd.it/7osxyipfikye1.png?width=1080&crop=smart&auto=webp&s=9f37932a3550ddd05a36c189059b3e3bcc091410', 'width': 1080}], 'source': {'height': 626, 'url': 'https://preview.redd.it/7osxyipfikye1.png?auto=webp&s=c373efd080ebd9267c2357ee9124f700f5b127c1', 'width': 1223}, 'variants': {}}]}
|
|||
Power efficient, affordable home server LLM hardware?
| 0 |
Hi all,
I've been running some small-ish LLMs as a coding assistant using llama.cpp & Tabby on my workstation laptop, and it's working pretty well!
My laptop has an Nvidia RTX A5000 with 16GB and it just about fits `Gemma3:12b-qat` as a chat / reasoning model and `Qwen2.5-coder:7b` for code completion (both using 4-bit quantization). They work well enough, and rather quickly, but it's impossible to use on battery or on my "on the go" older subnotebook.
I've been looking at options for a home server for running LLMs. I would prefer something at least as fast as the A5000, but I would also like to use (or at least try) a few bigger models. Gemma3:27b seems to provide significantly better results, and I'm keen to try the new Qwen3 models.
Power costs about 40 cents / kWh here, so power efficiency is important to me. The A5000 consumes about 35-50W when doing inference work and outputs about 37 tokens/sec for the 12b gemma3 model, so anything that exceeds that is fine, faster is obviously better.
Also it should run on Linux, so Apple silicon is unfortunately out of the question (I've tried running llama.cpp on Asahi Linux on an M2 Pro before using the Vulkan backend, and performance is pretty bad as it stands).
| 2025-05-03T13:24:29 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdswby/power_efficient_affordable_home_server_llm/
|
spaceman_
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdswby
| false | null |
t3_1kdswby
|
/r/LocalLLaMA/comments/1kdswby/power_efficient_affordable_home_server_llm/
| false | false |
self
| 0 | null |
Enable/Disable Reasoning Qwen 3
| 2 |
Is there a way we can turn on/off the reasoning mode either with a `llama-server` parameter or Open WebUI toggle?
| 2025-05-03T13:33:16 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdt2yb/enabledisable_reasoning_qwen_3/
|
Remarkable_Art5653
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdt2yb
| false | null |
t3_1kdt2yb
|
/r/LocalLLaMA/comments/1kdt2yb/enabledisable_reasoning_qwen_3/
| false | false |
self
| 2 | null |
New to AI stuff
| 11 |
Hello everyone.
My rig is: 4070 12GB + 32gb RAM
I just got into locally running my AI. I had a successfull run yesterday running in wsl ollama + gemma3:12B + openwebui. I wanted to ask how are you guys running your AI models, what are you using?
My end goal would be a chatbot in telegram that i could give tasks to over the internet, like : scrape this site, analyze this excel file locally. I would like to give it basically a folder on my pc that i would dump text files into for context. Is this possible?
Thank you for the time involved in reading this. Please excuse me for noob language.
PS: any informations given will be read.
| 2025-05-03T13:34:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdt43o/new_to_ai_stuff/
|
GIGKES
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdt43o
| false | null |
t3_1kdt43o
|
/r/LocalLLaMA/comments/1kdt43o/new_to_ai_stuff/
| false | false |
self
| 11 | null |
The significance of such a small model like qwen3-0.6B for mobile devices is immense.
| 1 |
[removed]
| 2025-05-03T13:42:00 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdt9cl/the_significance_of_such_a_small_model_like/
|
Aware-Ad-481
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdt9cl
| false | null |
t3_1kdt9cl
|
/r/LocalLLaMA/comments/1kdt9cl/the_significance_of_such_a_small_model_like/
| false | false |
self
| 1 | null |
Bought 3090, need emotional support
| 1 |
[removed]
| 2025-05-03T13:42:40 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdt9uv/bought_3090_need_emotional_support/
|
HandsOnDyk
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdt9uv
| false | null |
t3_1kdt9uv
|
/r/LocalLLaMA/comments/1kdt9uv/bought_3090_need_emotional_support/
| false | false |
self
| 1 | null |
glm-4-32b-0414 Aider Polyglot benchmark (scored 10%)
| 8 |
Hey everyone,
I recently conducted a benchmark on the GLM-4-32B-0414 model using aider polyglot and wanted to share my findings:
- dirname: 2025-05-02-18-07-24--NewHope
test_cases: 225
model: lm_studio/glm-4-32b-0414
edit_format: whole
commit_hash: e205629-dirty
pass_rate_1: 4.4
pass_rate_2: 10.2
pass_num_1: 10
pass_num_2: 23
percent_cases_well_formed: 99.1
error_outputs: 2
num_malformed_responses: 2
num_with_malformed_responses: 2
user_asks: 134
lazy_comments: 0
syntax_errors: 0
indentation_errors: 0
exhausted_context_windows: 0
test_timeouts: 3
total_tests: 225
command: aider --model lm_studio/glm-4-32b-0414
date: 2025-05-02
versions: 0.82.3.dev
seconds_per_case: 49.2
total_cost: 0.0000
**Only 10%**. Quite small I would say...
I experimented with different temperatures (0 and 0.8) and edit formats (whole vs. diff), but the results remained consistent. The low pass rates were unexpected, especially given the model's reported performance in other benchmarks and just the overall hype.
One potential factor could be the context window limitation of 32k tokens, which might have led to some malformed requests.
Has anyone else benchmarked this model or encountered similar results? I'd appreciate any insights or suggestions.
btw here is the command for the testing suite, if you had set this up using lm studio:
`LM_STUDIO_API_BASE=http://192.168.0.131:1234/v1 LM_STUDIO_API_KEY=dummy python3 benchmark/benchmark.py "NewHope" --model lm_studio/glm-4-32b-0414 --new --tries 2 --threads 1`
and you would need to create this entry in model-settings.yml :
- name: lm_studio/glm-4-32b-0414
use_temperature: 0.8
edit_format: whole
extra_params:
max_tokens: 32768
| 2025-05-03T14:07:56 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdtsx3/glm432b0414_aider_polyglot_benchmark_scored_10/
|
ryseek
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdtsx3
| false | null |
t3_1kdtsx3
|
/r/LocalLLaMA/comments/1kdtsx3/glm432b0414_aider_polyglot_benchmark_scored_10/
| false | false |
self
| 8 | null |
Incredible Maverick speeds on single RTX3090 - Ik_llama solved my issue
| 49 |
I was getting good generation speeds on Maverick before, but PP was slow.
This is now solved, I'm getting full GPU level performance on a 400B model with 1 gpu.
And the new Xeon DDR5 build takes it to the next level:
Xeon Platinum 8480 **ES** \- $170
8x 32GB DDR5 4800 RDIMM used - $722
1x Gigabyte MS03-CE0 - $753 (I got a MS73-HB1 but would recommend single CPU)
RTX 3090 - \~$750
Heatsink + PSU + Case + SSD = \~$500
prompt eval time = 835.47 ms / 372 tokens ( 2.25 ms per token, **445.26 tokens per second**
generation eval time = 43317.29 ms / 1763 runs ( 24.57 ms per token, **40.70 tokens per second**
prompt eval time = 3290.21 ms / 1623 tokens ( 2.03 ms per token, **493.28 tokens per second**
generation eval time = 7530.90 ms / 303 runs ( 24.85 ms per token, **40.23 tokens per second**
prompt eval time = 13713.39 ms / 7012 tokens ( 1.96 ms per token, **511.33 tokens per second**
generation eval time = 16773.69 ms / 584 runs ( 28.72 ms per token, **34.82 tokens per second**
This is with Ik\_Llama and the following command:
./llama-server -m Llama-4-Maverick-17B-128E-Instruct-UD-IQ4\_XS-00001-of-00005.gguf -c 32000 -fa -fmoe -amb 512 -rtr -ctk q8\_0 -ctv q8\_0 --host [0.0.0.0](http://0.0.0.0) \--port 8000 --alias Llama4-Maverick -ngl 99 -t 54 -ot ".\*ffn\_.\*\_exps.\*=CPU"
Using an ES cpu is somewhat risky, but a real 8480 cost $9k
This also works fine with an even cheaper DDR4 epyc cpu, getting 200+ Promp speeds and more like 20T/s gen with the same command.
This really makes me really hopeful for Llama 4 reasoner!
| 2025-05-03T14:44:09 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdul92/incredible_maverick_speeds_on_single_rtx3090_ik/
|
Conscious_Cut_6144
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdul92
| false | null |
t3_1kdul92
|
/r/LocalLLaMA/comments/1kdul92/incredible_maverick_speeds_on_single_rtx3090_ik/
| false | false |
self
| 49 | null |
deepseek r2 distill qwen 3?
| 37 |
hmm i really hope they make somehthing like that when the R2 comeout and that the community can push doing something like that what you think about this dream?
| 2025-05-03T14:55:05 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdutys/deepseek_r2_distill_qwen_3/
|
allforyi_mf
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdutys
| false | null |
t3_1kdutys
|
/r/LocalLLaMA/comments/1kdutys/deepseek_r2_distill_qwen_3/
| false | false |
self
| 37 | null |
Guide on Finetuning Reasoning Models
| 1 |
[removed]
| 2025-05-03T15:11:21 |
https://www.datacamp.com/tutorial/fine-tuning-deepseek-r1-reasoning-model
|
Infamous_Jaguar_2151
|
datacamp.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdv713
| false | null |
t3_1kdv713
|
/r/LocalLLaMA/comments/1kdv713/guide_on_finetuning_reasoning_models/
| false | false |
default
| 1 | null |
Is GLM-4's Long Context Performance Enough? An Undereducated Investigation
| 20 | 2025-05-03T15:12:59 |
https://adamniederer.com/blog/llm-context-benchmarks.html
|
vvimpcrvsh
|
adamniederer.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdv8by
| false | null |
t3_1kdv8by
|
/r/LocalLLaMA/comments/1kdv8by/is_glm4s_long_context_performance_enough_an/
| false | false |
default
| 20 | null |
|
Train a small language model to extract structured JSON from OCR text based on 'any user-defined schema'.
| 3 |
How would you guys proceed? So basically user can define any schema for example:
{
"invoice_no":"string",
"issued_to": {
"name": "string",
"address": "string" // Address of the client
},
"pay_to": {
"bank_name": "string", // Name of the bank
"name": "string", // Name
"account_no": "number"
},
"items":[
{
"description": "string",
"quantity": "number",
"unit_price": "number",
"total":"number"
}
],
"subtotal":"number",
"total":"number"
}
and we should get a response:
{
"invoice_no": "01234",
"issued_to": {
"name": "Richard Sanchez",
"address": "123 Anywhere St., Any City."
},
"pay_to": {
"bank_name": "Borcele Bank",
"name": "Adeline Palmerston",
"account_no": 012345678901
},
"items": [
{
"description": "Brand consultation",
"quantity": 1,
"unit_price": 100,
"total": 100
},
{
"description": "logo design",
"quantity": 1,
"unit_price": 100,
"total": 100
},
{
"description": "Website design",
"quantity": 1,
"unit_price": 100,
"total": 100
},
{
"description": "Social media templates",
"quantity": 1,
"unit_price": 100,
"total": 100
},
{
"description": "Brand photography",
"quantity": 1,
"unit_price": 100,
"total": 100
},
{
"description": "Brand guide",
"quantity": 1,
"unit_price": 100,
"total": 100
}
],
"subtotal": 400,
"total": 440
}
we will provide invoice text as context. Do you train a small mmodel(0.5B or 1.5B)? I can't send data online. I did try something and got some decent results. I will share that but before that I would like to know how you would try so i get unbiased opinions and see if I can improve..
| 2025-05-03T15:22:36 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdvg0j/train_a_small_language_model_to_extract/
|
maylad31
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdvg0j
| false | null |
t3_1kdvg0j
|
/r/LocalLLaMA/comments/1kdvg0j/train_a_small_language_model_to_extract/
| false | false |
self
| 3 | null |
Note to LLM researchers: we need graded benchmarks measuring levels of difficulty where models work at 100% accuracy
| 18 |
Just about all benchmarks I've seen are designed to be challenging, with no model reaching 100% accurate results, the main purpose being relative assessment of models against each other. In production use, however, there are situations where we need to know that for the given use case, the model we want to use will be 100% reliable and accurate. So we need benchmarks with different levels of difficulty, with the easiest levels reliably saturated by the smallest models, and onward from there. If we had this, it would take a lot of the guesswork out of our attempts to use small models for tasks that have to be done right 100% of the time.
Now I might be told that this is simply not possible, that no matter how easy a task, no LLM can be guaranteed to always produce 100% accurate output. I don't know if this is true, but even if it is, it could be accounted for and the small possibility of error accepted. As long as a reasonably thorough benchmark at a set level of difficutly results in 100%, that would be good enough, never mind that such perfection may not be attainable in production.
What do you all think? Would this be of use to you?
| 2025-05-03T15:29:59 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdvlv6/note_to_llm_researchers_we_need_graded_benchmarks/
|
ethereel1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdvlv6
| false | null |
t3_1kdvlv6
|
/r/LocalLLaMA/comments/1kdvlv6/note_to_llm_researchers_we_need_graded_benchmarks/
| false | false |
self
| 18 | null |
The GPT-4o sycophancy saga seems to be a case against open-source decentralized models?
| 0 |
Correct me if I am wrong, but it seems to me that much of the damage in this case could be mitigated because GPT-4o was a closed-source centralized model? One rollback and boom, no one on earth has access to it anymore. If a dangerously misaligned and powerful open source model was released like that, it would never be erased from public domain. Some providers/users would still be serving it to unsuspecting users/using it themselves either by mistake or due to malicious intent. What are the safeguards in place to prevent something like that from happening? This seems to me completely different case from open source programs, which allow anyone to inspect it under the hood and find out defects or malware (for e.g. the famous xz backdoor). There isn't anyway to do that (at present) for open weight models.
| 2025-05-03T15:36:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdvrjr/the_gpt4o_sycophancy_saga_seems_to_be_a_case/
|
obvithrowaway34434
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdvrjr
| false | null |
t3_1kdvrjr
|
/r/LocalLLaMA/comments/1kdvrjr/the_gpt4o_sycophancy_saga_seems_to_be_a_case/
| false | false |
self
| 0 | null |
Deepseek R2, when?
| 0 |
When do people think deepseek R2 will come out?
| 2025-05-03T16:13:37 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdwlp9/deepseek_r2_when/
|
MrMrsPotts
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdwlp9
| false | null |
t3_1kdwlp9
|
/r/LocalLLaMA/comments/1kdwlp9/deepseek_r2_when/
| false | false |
self
| 0 | null |
Qwen3 Coder incoming?
| 1 |
[removed]
| 2025-05-03T16:14:12 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdwm5y/qwen3_coder_incoming/
|
SandBlaster2000AD
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdwm5y
| false | null |
t3_1kdwm5y
|
/r/LocalLLaMA/comments/1kdwm5y/qwen3_coder_incoming/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'C_P6uz9UVliqHlGQQ8USWBUxafaMel9UdDuCfQaAAAA', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/oYLZZyhofxOu1J_YSi0tNbokcnHyMZqJ6IjLu8lpksc.jpg?width=108&crop=smart&auto=webp&s=f0a53ccca28e242797b207fe72c40e56700cdfd5', 'width': 108}], 'source': {'height': 200, 'url': 'https://external-preview.redd.it/oYLZZyhofxOu1J_YSi0tNbokcnHyMZqJ6IjLu8lpksc.jpg?auto=webp&s=2390e715787dc3dfadbf119a4945bbe5033f7282', 'width': 200}, 'variants': {}}]}
|
UI-Tars-1.5 reasoning never fails to entertain me.
| 1 |
[removed]
| 2025-05-03T16:35:04 |
Middle_Flow_2270
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdx2mm
| false | null |
t3_1kdx2mm
|
/r/LocalLLaMA/comments/1kdx2mm/uitars15_reasoning_never_fails_to_entertain_me/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'jtEWseW4a7pYyQm0HB8j0hcrnFfOs2Djsjf3M8TmlRc', 'resolutions': [{'height': 69, 'url': 'https://preview.redd.it/snyv99t1hlye1.jpeg?width=108&crop=smart&auto=webp&s=7dfe11a2a2874e2cb895dab2f6bfbf0aea24d703', 'width': 108}, {'height': 138, 'url': 'https://preview.redd.it/snyv99t1hlye1.jpeg?width=216&crop=smart&auto=webp&s=aea8d5984d822595ec30dd486a4df464501dca07', 'width': 216}, {'height': 204, 'url': 'https://preview.redd.it/snyv99t1hlye1.jpeg?width=320&crop=smart&auto=webp&s=d8ca7b6fde91029f3af09b5d4c6c83c480b96f57', 'width': 320}, {'height': 409, 'url': 'https://preview.redd.it/snyv99t1hlye1.jpeg?width=640&crop=smart&auto=webp&s=2d7ef807130ec56c28f8f55a30a1a872d64f12fe', 'width': 640}], 'source': {'height': 466, 'url': 'https://preview.redd.it/snyv99t1hlye1.jpeg?auto=webp&s=4fb3c544ca13890588501b92cdd83312f5b50b4b', 'width': 729}, 'variants': {}}]}
|
||
zero dollars vibe debugging menace
| 98 |
Been tweaking on building **Cloi** its *local* debugging agent that runs in your terminal. got sick of cloud models bleeding my wallet dry (o3 at $0.30 per request?? claude 3.7 still taking $0.05 a pop) so built something with zero dollar sign vibes.
the tech is straightforward: cloi deadass catches your error tracebacks, spins up your local LLM (phi/qwen/llama), and only with permission (we respectin boundaries), drops clean af patches directly to your files.
zero api key nonsense, no cloud tax - just pure on-device cooking with the models y'all are already optimizing FRFR
been working on this during my research downtime. If anyone's interested in exploring the implementation or wants to issue feedback: [https://github.com/cloi-ai/cloi](https://github.com/cloi-ai/cloi)
| 2025-05-03T16:45:56 |
AntelopeEntire9191
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdxbd7
| false | null |
t3_1kdxbd7
|
/r/LocalLLaMA/comments/1kdxbd7/zero_dollars_vibe_debugging_menace/
| false | false | 98 |
{'enabled': True, 'images': [{'id': 'b8GE0iJeZyULj_IkQcD00aH863qpsjPKqz4jEDwWbbU', 'resolutions': [{'height': 69, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=108&crop=smart&format=png8&s=c11ede4902dde5fe5f2876504b35aebc3aa9d0b1', 'width': 108}, {'height': 139, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=216&crop=smart&format=png8&s=12aab54e53e38d34db9d4e1511238929262e82a0', 'width': 216}, {'height': 207, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=320&crop=smart&format=png8&s=9571e4e8d62f4f8cb98b859ee6c75d068649f5c7', 'width': 320}, {'height': 414, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=640&crop=smart&format=png8&s=37356542f80cba5729c7207f13ed5ddbaeed8ce1', 'width': 640}, {'height': 621, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=960&crop=smart&format=png8&s=6d50ce647c4e2f064705ef1dde09d558ece4081c', 'width': 960}, {'height': 699, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=1080&crop=smart&format=png8&s=6da150fa75aeb7868cade4a149db57a40a93bf78', 'width': 1080}], 'source': {'height': 738, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?format=png8&s=6c4317e0d2f93f972dd3a29a2e731ac77017556a', 'width': 1140}, 'variants': {'gif': {'resolutions': [{'height': 69, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=108&crop=smart&s=b36665bb0cf1b127ad3adafb51cebc8f3bf24446', 'width': 108}, {'height': 139, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=216&crop=smart&s=22b6e51ac8efbf67c780d4ef90f46f007309341b', 'width': 216}, {'height': 207, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=320&crop=smart&s=6b3a9ee541dd6c8639779ff16f975bdd3327c318', 'width': 320}, {'height': 414, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=640&crop=smart&s=ed09e370865a1d4f607b60e7a14d5cce5bdb0ae9', 'width': 640}, {'height': 621, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=960&crop=smart&s=4789cb2931c63112ed675eb591bc0f0ad5725459', 'width': 960}, {'height': 699, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=1080&crop=smart&s=f41c0f3892b14bba46261e9641cd6a684f4f788c', 'width': 1080}], 'source': {'height': 738, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?s=072c69d5159d744dbc10105f459890483a1baea5', 'width': 1140}}, 'mp4': {'resolutions': [{'height': 69, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=108&format=mp4&s=570a2e145e23e101df131fa08fcf58df0d5224e4', 'width': 108}, {'height': 139, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=216&format=mp4&s=d97c4acb06fcd96a967693273da0546784078827', 'width': 216}, {'height': 207, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=320&format=mp4&s=9c04bba4eb24573ecf9ae56ba4ea381a14c3ffaa', 'width': 320}, {'height': 414, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=640&format=mp4&s=4c5165c02983a1a3dc872f197ad8541368205edf', 'width': 640}, {'height': 621, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=960&format=mp4&s=02f77ebd013fd7981ee558c28e2e1aa5be935cb9', 'width': 960}, {'height': 699, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?width=1080&format=mp4&s=6965f579a1769299a8572a821a1bb1386d879cff', 'width': 1080}], 'source': {'height': 738, 'url': 'https://preview.redd.it/wckdwzhiilye1.gif?format=mp4&s=e7a74e686cbf7461879bd9a5e917f66a29970b9b', 'width': 1140}}}}]}
|
||
Cómo entrenar un loRA realista con pony en seaart?
| 1 |
[removed]
| 2025-05-03T16:54:27 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdxi8m/cómo_entrenar_un_lora_realista_con_pony_en_seaart/
|
Quiet_Dot_1743
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdxi8m
| false | null |
t3_1kdxi8m
|
/r/LocalLLaMA/comments/1kdxi8m/cómo_entrenar_un_lora_realista_con_pony_en_seaart/
| false | false |
self
| 1 | null |
CLAUDE MAX or Augmentcode?
| 0 |
Hi guys I want to make my own indie game and I am done with all the basics of unity. I have coded around 1000 lines - a solid base for my game. I have the main menu done, dialogue system and a audio manager.
I used claude pro to do this. And it's just so solid.
I want to invest in Claude code (200 usd Max subscription ) . But can I get it to work on windows and in Unity somehow?
Or is Augment code better?
For 30 dollars I saw UNLIMITED ACCESS.
Please help me.
| 2025-05-03T17:12:21 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdxwxp/claude_max_or_augmentcode/
|
backinthe90siwasinav
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdxwxp
| false | null |
t3_1kdxwxp
|
/r/LocalLLaMA/comments/1kdxwxp/claude_max_or_augmentcode/
| false | false |
self
| 0 | null |
How to use LLMs to help build eval datasets for your RAG system
| 1 |
[removed]
| 2025-05-03T17:17:46 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdy1g7/how_to_use_llms_to_help_build_eval_datasets_for/
|
zjost85
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdy1g7
| false | null |
t3_1kdy1g7
|
/r/LocalLLaMA/comments/1kdy1g7/how_to_use_llms_to_help_build_eval_datasets_for/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'QuL27Zy3dXZqolE83CmgaLqydP4_TwzPqT9iH27RuTw', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/EvjYIGZqXiEGGZtnzv6zKOcaGdPLm-YXXNXN1sKtGpY.jpg?width=108&crop=smart&auto=webp&s=3c1a8ea2fc9059078491467e5cb3e7738337fded', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/EvjYIGZqXiEGGZtnzv6zKOcaGdPLm-YXXNXN1sKtGpY.jpg?width=216&crop=smart&auto=webp&s=305ba44075511453847df9e7473edc24b85b6536', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/EvjYIGZqXiEGGZtnzv6zKOcaGdPLm-YXXNXN1sKtGpY.jpg?width=320&crop=smart&auto=webp&s=a503e403467649f118507216a97cf800a80366fa', 'width': 320}], 'source': {'height': 270, 'url': 'https://external-preview.redd.it/EvjYIGZqXiEGGZtnzv6zKOcaGdPLm-YXXNXN1sKtGpY.jpg?auto=webp&s=100fbe03eabc1d287aadd36cc4320509784238f3', 'width': 480}, 'variants': {}}]}
|
Microsoft is cooking coding models, NextCoder.
| 264 | 2025-05-03T17:26:07 |
https://huggingface.co/collections/microsoft/nextcoder-6815ee6bfcf4e42f20d45028
|
mlon_eusk-_-
|
huggingface.co
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdy8ia
| false | null |
t3_1kdy8ia
|
/r/LocalLLaMA/comments/1kdy8ia/microsoft_is_cooking_coding_models_nextcoder/
| false | false | 264 |
{'enabled': False, 'images': [{'id': '7_jhFTazab6GMtEoANxssbRBy-NQcSp84SYt3Tyoa40', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/Ucgd876uHUi68EnzSBM_dLw0RYz0XUiaeONwhBsyglM.jpg?width=108&crop=smart&auto=webp&s=d96a575382e231e9327b16721a0ac91ec9062b6f', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/Ucgd876uHUi68EnzSBM_dLw0RYz0XUiaeONwhBsyglM.jpg?width=216&crop=smart&auto=webp&s=f6519c23ca20ae351174f2ef2bec2a54b4b4cf76', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/Ucgd876uHUi68EnzSBM_dLw0RYz0XUiaeONwhBsyglM.jpg?width=320&crop=smart&auto=webp&s=2f270078216f0b3455ab5d472a6eaf313052e495', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/Ucgd876uHUi68EnzSBM_dLw0RYz0XUiaeONwhBsyglM.jpg?width=640&crop=smart&auto=webp&s=7d72cbc224e62aa30d0164faecb579c9a6eee406', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/Ucgd876uHUi68EnzSBM_dLw0RYz0XUiaeONwhBsyglM.jpg?width=960&crop=smart&auto=webp&s=9ef5709f98a00d2bad049dd2cd47b1955026548a', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/Ucgd876uHUi68EnzSBM_dLw0RYz0XUiaeONwhBsyglM.jpg?width=1080&crop=smart&auto=webp&s=d2769e30559e7f04ab9a1dad6894f46fc3965228', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/Ucgd876uHUi68EnzSBM_dLw0RYz0XUiaeONwhBsyglM.jpg?auto=webp&s=62fcd09a8a31c34f6a1a8ceacd64cd05a54fb15b', 'width': 1200}, 'variants': {}}]}
|
||
With the rise of LLMs with optional CoT, when to use it?
| 1 |
[removed]
| 2025-05-03T17:28:39 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdyakb/with_the_rise_of_llms_with_optional_cot_when_to/
|
pixelar-cat
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdyakb
| false | null |
t3_1kdyakb
|
/r/LocalLLaMA/comments/1kdyakb/with_the_rise_of_llms_with_optional_cot_when_to/
| false | false |
self
| 1 | null |
Hey step-bro, that's HF forum, not the AI chat...
| 392 | 2025-05-03T17:28:57 |
Cool-Chemical-5629
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdyasr
| false | null |
t3_1kdyasr
|
/r/LocalLLaMA/comments/1kdyasr/hey_stepbro_thats_hf_forum_not_the_ai_chat/
| false | false | 392 |
{'enabled': True, 'images': [{'id': 'pntRmVcOnztYdt8CI8SDyvJaX15z94CYpg1_xWIKVXs', 'resolutions': [{'height': 105, 'url': 'https://preview.redd.it/3a4xy047qlye1.png?width=108&crop=smart&auto=webp&s=20cd816332feeaf631e1bbbd06e68754282bfe40', 'width': 108}, {'height': 210, 'url': 'https://preview.redd.it/3a4xy047qlye1.png?width=216&crop=smart&auto=webp&s=0e68dca8b2d704ec6cc25ff4df998351699a5b04', 'width': 216}, {'height': 312, 'url': 'https://preview.redd.it/3a4xy047qlye1.png?width=320&crop=smart&auto=webp&s=46d770c0b9baa33161861c7eaea174f7b8ab0d3a', 'width': 320}, {'height': 624, 'url': 'https://preview.redd.it/3a4xy047qlye1.png?width=640&crop=smart&auto=webp&s=2f0827b40d5dad9a38133015ced4dd0b976f205e', 'width': 640}], 'source': {'height': 903, 'url': 'https://preview.redd.it/3a4xy047qlye1.png?auto=webp&s=e15de4fa497501047d40ee3ba515cf4ecede57de', 'width': 926}, 'variants': {}}]}
|
|||
Chatterui and local models
| 4 |
Hello lads,
I wanted to try some models offline on my smartphone so I installed Chatterui and downloaded various ggufs both from 1b and 660m (gemma3, quen3 and others) but as soon as it starts to load the model the application closes.
Am I doing something wrong? Do you have suggestions?
Thank you all
Xiaomi redmi note 12 pro with 8gb of Ram
| 2025-05-03T17:30:33 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdyc4z/chatterui_and_local_models/
|
carstele69
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdyc4z
| false | null |
t3_1kdyc4z
|
/r/LocalLLaMA/comments/1kdyc4z/chatterui_and_local_models/
| false | false |
self
| 4 | null |
Need clarification on Qwen3-30B-a3b-q8 and Qwen3-4b-q8 performance and use cases.
| 1 |
I have a framework that switches between chat mode and analysis mode and runs both on Ollama 0.6.6, loading each model separately as needed. These modes are run by two separate models because I haven't added support for hybrid models yet so I have to load them separately for now.
For Chat Mode, I use Q3-4b-q8 - /no_think - 12k context length
For Analysis Mode, I use Q3-30b-a3b - /think - 12k context length
The problem is that I have a prompt that has a very complicated set of instructions containing a lot of input from many different sources converted into text (images, audio, etc.).
Normally larger models (14b and higher) handle this well and smaller models struggle, which is expected.
However, in chat mode, it seems that Q3-4b consistently handles this much better than the a3b model while both handle themselves well in Analysis mode when their thinking is enabled.
In Chat mode, a3b struggles a lot, usually giving me blank responses if the conversation history is around 9K tokens long.
I know it performs better in analysis mode, but I would like to test it out in Chat mode because I assumed that even with /no_think it would blow the 4b model out of the water but in reality the exact opposite is happening.
Considering its a MoE model, is it possible that I'm pushing the model too far with the complexity of the instructions? My hypothesis is that the MoE is supposed to handle requests that require precision or specialization, which is why it gives me concrete answers with /think enabled but it gives me blank responses in long context with /no_think.
| 2025-05-03T17:35:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdyg3f/need_clarification_on_qwen330ba3bq8_and_qwen34bq8/
|
swagonflyyyy
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdyg3f
| false | null |
t3_1kdyg3f
|
/r/LocalLLaMA/comments/1kdyg3f/need_clarification_on_qwen330ba3bq8_and_qwen34bq8/
| false | false |
self
| 1 | null |
How can I "inject" new data into an LLM? And which LLM would be best for me?
| 7 |
How can I "inject" new data into an LLM? And which LLM would be best for me?
I'm not talking about adding a document to the chat, but rather integrating, for example, a number of books and having them... "thought out."
Let's say I'm reading a relatively modern philosophy author and the LLM I'm using doesn't know much about it. Can I add all the author's books I have in .txt format? Do I need a high-capacity LLM to understand them, or is it not necessary? Perhaps a low-capacity LLM can still understand them if it has all the books?
But can this still be done?
I think it's called fine-tuning... would it take a long time on an 8GB RAM and 32GB RAM machine?
| 2025-05-03T17:54:49 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdyw3q/how_can_i_inject_new_data_into_an_llm_and_which/
|
9acca9
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdyw3q
| false | null |
t3_1kdyw3q
|
/r/LocalLLaMA/comments/1kdyw3q/how_can_i_inject_new_data_into_an_llm_and_which/
| false | false |
self
| 7 | null |
Any good pdf/epub to TTS audiobook generators?
| 1 |
[removed]
| 2025-05-03T18:17:41 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdzfj1/any_good_pdfepub_to_tts_audiobook_generators/
|
Virtamancer
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdzfj1
| false | null |
t3_1kdzfj1
|
/r/LocalLLaMA/comments/1kdzfj1/any_good_pdfepub_to_tts_audiobook_generators/
| false | false |
self
| 1 | null |
Handling Sensitive Data (PII) in Local RAG Pipelines?
| 1 |
[removed]
| 2025-05-03T18:38:30 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdzwsv/handling_sensitive_data_pii_in_local_rag_pipelines/
|
Revolutionary-Bet-58
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdzwsv
| false | null |
t3_1kdzwsv
|
/r/LocalLLaMA/comments/1kdzwsv/handling_sensitive_data_pii_in_local_rag_pipelines/
| false | false |
self
| 1 | null |
Qwen3 on 3060 12GB VRAM and 16GB RAM
| 9 |
is there any way to run this LLM on my PC? how to install and which model is suitable for my PC?
| 2025-05-03T18:38:40 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdzwyf/qwen3_on_3060_12gb_vram_and_16gb_ram/
|
RaviieR
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdzwyf
| false | null |
t3_1kdzwyf
|
/r/LocalLLaMA/comments/1kdzwyf/qwen3_on_3060_12gb_vram_and_16gb_ram/
| false | false |
self
| 9 | null |
Has there been a project that recreates OpenAI's tool-assisted image gen?
| 3 |
Pure curiosity, and because I couldn't find anything about this.
I have been using OpenAI's image gen tool to help me with better visual understanding of stuff due to my visual impairment by either recreating certain things in large or from different angles. And while not perfect, it helps a whole lot.
But, I would love to see this run locally - if possible, that is. So I tried to find such a tool-assisted image gen to run locally; but other than the "classics" like InvokeAI, automatic1111 and stablediffusion.cpp, I couldn't find anything that actually does what OpenAI currently does.
Perhaps I overlooked something. Hence, I wanted to ask if you perhaps know of something?
Thanks!
| 2025-05-03T18:40:06 |
https://www.reddit.com/r/LocalLLaMA/comments/1kdzy5l/has_there_been_a_project_that_recreates_openais/
|
IngwiePhoenix
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kdzy5l
| false | null |
t3_1kdzy5l
|
/r/LocalLLaMA/comments/1kdzy5l/has_there_been_a_project_that_recreates_openais/
| false | false |
self
| 3 | null |
Are there actually any good "long-form" models for local use?
| 1 |
[removed]
| 2025-05-03T18:48:26 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke050n/are_there_actually_any_good_longform_models_for/
|
Barefooter1234
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke050n
| false | null |
t3_1ke050n
|
/r/LocalLLaMA/comments/1ke050n/are_there_actually_any_good_longform_models_for/
| false | false |
self
| 1 | null |
Anticipating a need for thinking in Qwen3
| 1 |
[removed]
| 2025-05-03T18:50:44 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke06xj/anticipating_a_need_for_thinking_in_qwen3/
|
Professional-Onion-7
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke06xj
| false | null |
t3_1ke06xj
|
/r/LocalLLaMA/comments/1ke06xj/anticipating_a_need_for_thinking_in_qwen3/
| false | false |
self
| 1 | null |
Anticipating a need for thinking in Qwen3
| 1 |
[removed]
| 2025-05-03T19:08:22 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke0ll9/anticipating_a_need_for_thinking_in_qwen3/
|
extraYogurtnote
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke0ll9
| false | null |
t3_1ke0ll9
|
/r/LocalLLaMA/comments/1ke0ll9/anticipating_a_need_for_thinking_in_qwen3/
| false | false |
self
| 1 | null |
Surprising results fine tuning Qwen3-4B
| 40 |
I’ve had a lot of experience fine tuning Qwen2.5 models on a proprietary programming language which wasn’t in pre-training data. I have an extensive SFT dataset which I’ve used with pretty decent success on the Qwen2.5 models.
Naturally when the latest Qwen3 crop dropped I was keen on seeing the results I’ll get with them.
Here’s the strange part:
I use an evaluation dataset of 50 coding tasks which I check against my fine tuned models. I actually send the model’s response to a compiler to check if it’s legible code.
Fine tuned Qwen3-4B (Default) Thinking ON - 40% success rate
Fine tuned Qwen3-4B Thinking OFF - 64% success rate
WTF? (Sorry for being crass)
A few side notes:
- These are both great results, base Qwen3-4B scores 0% and they are much better than Qwen2.5-3B
- My SFT dataset does not contain <think>ing tags
- I’m doing a full parameter fine tune at BF16 precision. No LoRA’s or quants.
Would love to hear some theories on why this is happening. And any ideas how to improve this.
As I said above, in general these models are awesome and performing (for my purposes) several factors better than Qwen2.5. Can’t wait to fine tune bigger sizes soon (as soon as I figure this out).
| 2025-05-03T20:02:10 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke1sei/surprising_results_fine_tuning_qwen34b/
|
indicava
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke1sei
| false | null |
t3_1ke1sei
|
/r/LocalLLaMA/comments/1ke1sei/surprising_results_fine_tuning_qwen34b/
| false | false |
self
| 40 | null |
Another Attempt to Measure Speed for Qwen3 MoE on 2x4090, 2x3090, M3 Max with Llama.cpp, VLLM, MLX
| 46 |
First, thank you all the people who gave constructive feedback on my previous attempt. Hopefully this is better. :)
### Observation
TL;TR: As expected, fastest to slowest: RTX 4090 VLLM, RTX 4090 Llama.CPP, RTX 3090 Llama.CPP, M3 Max MLX, M3 Max Llama.CPP
M3 Max was slow but surprisingly not too terrible when using this particular model with MLX.
### Notes
To ensure consistency, I used a custom Python script that sends requests to the server via the OpenAI-compatible API. Metrics were calculated as follows:
* Time to First Token (TTFT): Measured from the start of the streaming request to the first streaming event received.
* Prompt Processing Speed (PP): Number of prompt tokens divided by TTFT.
* Token Generation Speed (TG): Number of generated tokens divided by (total duration - TTFT).
The displayed results were truncated to two decimal places, but the calculations used full precision.
Some engines, like MLX, don't let you disable prompt caching. To work around this, I made the script to prepend 40% new material in the beginning of next longer prompt to avoid caching effect.
### Setup
* VLLM 0.8.5.post1
* MLX-LM 0.24.0, MLX 0.25.1
* Llama.CPP 5269
Each row in the results represents a test (a specific combination of machine, engine, and prompt length). There are 5 tests per prompt length.
* Setup 1: 2xRTX-4090, VLLM, FP8, tensor-parallel-size 2
* Setup 2: 2xRTX-4090, Llama.cpp, q8_0, flash attention
* Setup 3: 2x3090, Llama.cpp, q8_0, flash attention
* Setup 4: M3Max, MLX, 8bit
* Setup 5: M3Max, Llama.cpp, q8_0, flash attention
VLLM doesn't support Mac. Also there's no test with RTX-3090 and VLLM either because you can't run Qwen3 MoE in FP8 with RTX-3090 using VLLM.
| Machine | Engine | Prompt Tokens | PP | TTFT | Generated Tokens | TG | Duration |
| ------- | ------ | ------------- | ----------------------- | ------------------- | ---------------- | ---------------------- | -------------- |
| rtx4090 | VLLM | 702 | 6823.88 | 0.10 | 1334 | 93.73 | 14.34 |
| RTX4090 | LCPP | 702 | 2521.87 | 0.28 | 1540 | 100.87 | 15.55 |
| RTX3090 | LCPP | 702 | 1632.82 | 0.43 | 1258 | 84.04 | 15.40 |
| M3Max | MLX | 702 | 1216.27 | 0.57 | 1296 | 65.69 | 20.30 |
| M3Max | LCPP | 702 | 290.22 | 2.42 | 1485 | 55.79 | 29.04 |
| rtx4090 | VLLM | 959 | 6837.26 | 0.14 | 1337 | 94.74 | 14.25 |
| RTX4090 | LCPP | 959 | 2657.34 | 0.36 | 1187 | 97.13 | 12.58 |
| RTX3090 | LCPP | 959 | 1685.90 | 0.57 | 1487 | 83.67 | 18.34 |
| M3Max | MLX | 959 | 1214.74 | 0.79 | 1523 | 65.09 | 24.18 |
| M3Max | LCPP | 959 | 465.91 | 2.06 | 1337 | 55.43 | 26.18 |
| rtx4090 | VLLM | 1306 | 7214.16 | 0.18 | 1167 | 94.17 | 12.57 |
| RTX4090 | LCPP | 1306 | 2646.48 | 0.49 | 1114 | 98.95 | 11.75 |
| RTX3090 | LCPP | 1306 | 1674.10 | 0.78 | 995 | 83.36 | 12.72 |
| M3Max | MLX | 1306 | 1258.91 | 1.04 | 1119 | 64.76 | 18.31 |
| M3Max | LCPP | 1306 | 458.79 | 2.85 | 1213 | 55.00 | 24.90 |
| rtx4090 | VLLM | 1774 | 7857.53 | 0.23 | 1353 | 93.24 | 14.74 |
| RTX4090 | LCPP | 1774 | 2625.51 | 0.68 | 1282 | 98.68 | 13.67 |
| RTX3090 | LCPP | 1774 | 1730.67 | 1.03 | 1411 | 82.66 | 18.09 |
| M3Max | MLX | 1774 | 1276.55 | 1.39 | 1330 | 63.03 | 22.49 |
| M3Max | LCPP | 1774 | 321.31 | 5.52 | 1281 | 54.26 | 29.13 |
| rtx4090 | VLLM | 2584 | 7851.00 | 0.33 | 1369 | 92.48 | 15.13 |
| RTX4090 | LCPP | 2584 | 2634.01 | 0.98 | 1308 | 97.20 | 14.44 |
| RTX3090 | LCPP | 2584 | 1728.13 | 1.50 | 1334 | 81.80 | 17.80 |
| M3Max | MLX | 2584 | 1302.66 | 1.98 | 1247 | 60.79 | 22.49 |
| M3Max | LCPP | 2584 | 449.35 | 5.75 | 1321 | 53.06 | 30.65 |
| rtx4090 | VLLM | 3557 | 8619.84 | 0.41 | 1682 | 92.46 | 18.60 |
| RTX4090 | LCPP | 3557 | 2684.50 | 1.33 | 2000 | 93.68 | 22.67 |
| RTX3090 | LCPP | 3557 | 1779.73 | 2.00 | 1414 | 80.31 | 19.60 |
| M3Max | MLX | 3557 | 1272.91 | 2.79 | 2001 | 59.81 | 36.25 |
| M3Max | LCPP | 3557 | 443.93 | 8.01 | 1481 | 51.52 | 36.76 |
| rtx4090 | VLLM | 4739 | 7944.01 | 0.60 | 1710 | 91.43 | 19.30 |
| RTX4090 | LCPP | 4739 | 2622.29 | 1.81 | 1082 | 91.46 | 13.64 |
| RTX3090 | LCPP | 4739 | 1736.44 | 2.73 | 1968 | 78.02 | 27.95 |
| M3Max | MLX | 4739 | 1239.93 | 3.82 | 1836 | 58.63 | 35.14 |
| M3Max | LCPP | 4739 | 421.45 | 11.24 | 1472 | 49.94 | 40.72 |
| rtx4090 | VLLM | 6520 | 8330.26 | 0.78 | 1588 | 90.54 | 18.32 |
| RTX4090 | LCPP | 6520 | 2616.54 | 2.49 | 1471 | 87.03 | 19.39 |
| RTX3090 | LCPP | 6520 | 1726.75 | 3.78 | 2000 | 75.44 | 30.29 |
| M3Max | MLX | 6520 | 1164.00 | 5.60 | 1546 | 55.89 | 33.26 |
| M3Max | LCPP | 6520 | 418.88 | 15.57 | 1998 | 47.61 | 57.53 |
| rtx4090 | VLLM | 9101 | 8156.34 | 1.12 | 1571 | 88.01 | 18.97 |
| RTX4090 | LCPP | 9101 | 2563.10 | 3.55 | 1342 | 83.52 | 19.62 |
| RTX3090 | LCPP | 9101 | 1661.47 | 5.48 | 1445 | 72.36 | 25.45 |
| M3Max | MLX | 9101 | 1061.38 | 8.57 | 1601 | 52.07 | 39.32 |
| M3Max | LCPP | 9101 | 397.69 | 22.88 | 1941 | 44.81 | 66.20 |
| rtx4090 | VLLM | 12430 | 6590.37 | 1.89 | 1805 | 84.48 | 23.25 |
| RTX4090 | LCPP | 12430 | 2441.21 | 5.09 | 1573 | 78.33 | 25.17 |
| RTX3090 | LCPP | 12430 | 1615.05 | 7.70 | 1150 | 68.79 | 24.41 |
| M3Max | MLX | 12430 | 954.98 | 13.01 | 1627 | 47.89 | 46.99 |
| M3Max | LCPP | 12430 | 359.69 | 34.56 | 1291 | 41.95 | 65.34 |
| rtx4090 | VLLM | 17078 | 6539.04 | 2.61 | 1230 | 83.61 | 17.32 |
| RTX4090 | LCPP | 17078 | 2362.40 | 7.23 | 1217 | 71.79 | 24.18 |
| RTX3090 | LCPP | 17078 | 1524.14 | 11.21 | 1229 | 65.38 | 30.00 |
| M3Max | MLX | 17078 | 829.37 | 20.59 | 2001 | 41.34 | 68.99 |
| M3Max | LCPP | 17078 | 330.01 | 51.75 | 1461 | 38.28 | 89.91 |
| rtx4090 | VLLM | 23658 | 6645.42 | 3.56 | 1310 | 81.88 | 19.56 |
| RTX4090 | LCPP | 23658 | 2225.83 | 10.63 | 1213 | 63.60 | 29.70 |
| RTX3090 | LCPP | 23658 | 1432.59 | 16.51 | 1058 | 60.61 | 33.97 |
| M3Max | MLX | 23658 | 699.38 | 33.82 | 2001 | 35.56 | 90.09 |
| M3Max | LCPP | 23658 | 294.29 | 80.39 | 1681 | 33.96 | 129.88 |
| rtx4090 | VLLM | 33525 | 5680.62 | 5.90 | 1138 | 77.42 | 20.60 |
| RTX4090 | LCPP | 33525 | 2051.73 | 16.34 | 990 | 54.96 | 34.35 |
| RTX3090 | LCPP | 33525 | 1287.74 | 26.03 | 1272 | 54.62 | 49.32 |
| M3Max | MLX | 33525 | 557.25 | 60.16 | 1328 | 28.26 | 107.16 |
| M3Max | LCPP | 33525 | 250.40 | 133.89 | 1453 | 29.17 | 183.69 |
| 2025-05-03T20:20:44 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke26sl/another_attempt_to_measure_speed_for_qwen3_moe_on/
|
chibop1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke26sl
| false | null |
t3_1ke26sl
|
/r/LocalLLaMA/comments/1ke26sl/another_attempt_to_measure_speed_for_qwen3_moe_on/
| false | false |
self
| 46 |
{'enabled': False, 'images': [{'id': 'vBAA2fS35uaiZy2nfiNIf6fASXQm_rW0ObMpevJyTp4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/5gpXdrU9UGIQ6qHsiNn-wsS_L2h19b1jjZMq_kBe0SU.jpg?width=108&crop=smart&auto=webp&s=6a3475eb6a52d991b0637bac936ba458d31c51c6', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/5gpXdrU9UGIQ6qHsiNn-wsS_L2h19b1jjZMq_kBe0SU.jpg?width=216&crop=smart&auto=webp&s=4d202c87cee404ab142051f42f8b1dd3a97f0ac4', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/5gpXdrU9UGIQ6qHsiNn-wsS_L2h19b1jjZMq_kBe0SU.jpg?width=320&crop=smart&auto=webp&s=526b95cd15ed26f3263ee1e7df3110d9724b9618', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/5gpXdrU9UGIQ6qHsiNn-wsS_L2h19b1jjZMq_kBe0SU.jpg?width=640&crop=smart&auto=webp&s=41e12c31898497a53a2b2480498db23848068127', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/5gpXdrU9UGIQ6qHsiNn-wsS_L2h19b1jjZMq_kBe0SU.jpg?width=960&crop=smart&auto=webp&s=c6dd8ff5cbbaee190c8acf621638fd016f8d7555', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/5gpXdrU9UGIQ6qHsiNn-wsS_L2h19b1jjZMq_kBe0SU.jpg?width=1080&crop=smart&auto=webp&s=6d3285736c9a2e3365b45f46371ccb0a1e1ba1b8', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/5gpXdrU9UGIQ6qHsiNn-wsS_L2h19b1jjZMq_kBe0SU.jpg?auto=webp&s=873220fc0959c32c7cb461483a88130b2583bf82', 'width': 1200}, 'variants': {}}]}
|
Dockerfile for Running BitNet-b1.58-2B-4T on ARM
| 14 |
###Repo
[GitHub: ajsween/bitnet-b1-58-arm-docker](https://github.com/ajsween/bitnet-b1-58-arm-docker)
I put this Dockerfile together so I could run the BitNet 1.58 model with less hassle on my M-series MacBook. Hopefully its useful to some else and saves you some time getting it running locally.
###Run interactive:
docker run -it --rm bitnet-b1.58-2b-4t-arm:latest
###Run noninteractive with arguments:
docker run --rm bitnet-b1.58-2b-4t-arm:latest \
-m models/BitNet-b1.58-2B-4T/ggml-model-i2_s.gguf \
-p "Hello from BitNet on Pi4!"
###Reference for run_interference.py (ENTRYPOINT):
usage: run_inference.py [-h] [-m MODEL] [-n N_PREDICT] -p PROMPT [-t THREADS] [-c CTX_SIZE] [-temp TEMPERATURE] [-cnv]
Run inference
optional arguments:
-h, --help show this help message and exit
-m MODEL, --model MODEL
Path to model file
-n N_PREDICT, --n-predict N_PREDICT
Number of tokens to predict when generating text
-p PROMPT, --prompt PROMPT
Prompt to generate text from
-t THREADS, --threads THREADS
Number of threads to use
-c CTX_SIZE, --ctx-size CTX_SIZE
Size of the prompt context
-temp TEMPERATURE, --temperature TEMPERATURE
Temperature, a hyperparameter that controls the randomness of the generated text
-cnv, --conversation Whether to enable chat mode or not (for instruct models.)
(When this option is turned on, the prompt specified by -p will be used as the system prompt.)
###Dockerfile
# Build stage
FROM python:3.9-slim AS builder
# Set environment variables
ENV DEBIAN_FRONTEND=noninteractive
ENV PYTHONDONTWRITEBYTECODE=1
ENV PYTHONUNBUFFERED=1
# Install build dependencies
RUN apt-get update && apt-get install -y \
python3-pip \
python3-dev \
cmake \
build-essential \
git \
software-properties-common \
wget \
&& rm -rf /var/lib/apt/lists/*
# Install LLVM
RUN wget -O - https://apt.llvm.org/llvm.sh | bash -s 18
# Clone the BitNet repository
WORKDIR /build
RUN git clone --recursive https://github.com/microsoft/BitNet.git
# Install Python dependencies
RUN pip install --no-cache-dir -r /build/BitNet/requirements.txt
# Build BitNet
WORKDIR /build/BitNet
RUN pip install --no-cache-dir -r requirements.txt \
&& python utils/codegen_tl1.py \
--model bitnet_b1_58-3B \
--BM 160,320,320 \
--BK 64,128,64 \
--bm 32,64,32 \
&& export CC=clang-18 CXX=clang++-18 \
&& mkdir -p build && cd build \
&& cmake .. -DCMAKE_BUILD_TYPE=Release \
&& make -j$(nproc)
# Download the model
RUN huggingface-cli download microsoft/BitNet-b1.58-2B-4T-gguf \
--local-dir /build/BitNet/models/BitNet-b1.58-2B-4T
# Convert the model to GGUF format and sets up env. Probably not needed.
RUN python setup_env.py -md /build/BitNet/models/BitNet-b1.58-2B-4T -q i2_s
# Final stage
FROM python:3.9-slim
# Set environment variables. All but the last two are not used as they don't expand in the CMD step.
ENV MODEL_PATH=/app/models/BitNet-b1.58-2B-4T/ggml-model-i2_s.gguf
ENV NUM_TOKENS=1024
ENV NUM_THREADS=4
ENV CONTEXT_SIZE=4096
ENV PROMPT="Hello from BitNet!"
ENV PYTHONUNBUFFERED=1
ENV LD_LIBRARY_PATH=/usr/local/lib
# Copy from builder stage
WORKDIR /app
COPY --from=builder /build/BitNet /app
# Install Python dependencies (only runtime)
RUN <<EOF
pip install --no-cache-dir -r /app/requirements.txt
cp /app/build/3rdparty/llama.cpp/ggml/src/libggml.so /usr/local/lib
cp /app/build/3rdparty/llama.cpp/src/libllama.so /usr/local/lib
EOF
# Set working directory
WORKDIR /app
# Set entrypoint for more flexibility
ENTRYPOINT ["python", "./run_inference.py"]
# Default command arguments
CMD ["-m", "/app/models/BitNet-b1.58-2B-4T/ggml-model-i2_s.gguf", "-n", "1024", "-cnv", "-t", "4", "-c", "4096", "-p", "Hello from BitNet!"]
| 2025-05-03T20:38:00 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke2kcp/dockerfile_for_running_bitnetb1582b4t_on_arm/
|
ajsween
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke2kcp
| false | null |
t3_1ke2kcp
|
/r/LocalLLaMA/comments/1ke2kcp/dockerfile_for_running_bitnetb1582b4t_on_arm/
| false | false |
self
| 14 |
{'enabled': False, 'images': [{'id': 'etNoPMX9wm3GDz9Ov7nWhcSrO2m2cKmHumPuST9ZEe8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/9cqW0OCcUeQISM2NOrxFqRit1tmxMKPpZMn6ujN60Q4.jpg?width=108&crop=smart&auto=webp&s=ee6493d52668501015c9c37727223af27bc58c19', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/9cqW0OCcUeQISM2NOrxFqRit1tmxMKPpZMn6ujN60Q4.jpg?width=216&crop=smart&auto=webp&s=e0a5d981b60ad8fa4d1b1e6f329ceef73a6f9f66', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/9cqW0OCcUeQISM2NOrxFqRit1tmxMKPpZMn6ujN60Q4.jpg?width=320&crop=smart&auto=webp&s=ed6fd576550e6298c07b25a1babb999f1efce5d7', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/9cqW0OCcUeQISM2NOrxFqRit1tmxMKPpZMn6ujN60Q4.jpg?width=640&crop=smart&auto=webp&s=36e0d7ef0fad30609e36aa65a6c24df741ff1465', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/9cqW0OCcUeQISM2NOrxFqRit1tmxMKPpZMn6ujN60Q4.jpg?width=960&crop=smart&auto=webp&s=ff589f2ed305815c720180dd1d939739c585ff62', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/9cqW0OCcUeQISM2NOrxFqRit1tmxMKPpZMn6ujN60Q4.jpg?width=1080&crop=smart&auto=webp&s=111625acf9a59716afc5d2f0695b491fe9db5773', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/9cqW0OCcUeQISM2NOrxFqRit1tmxMKPpZMn6ujN60Q4.jpg?auto=webp&s=c93bc9ded4960cbd12ccadab5fea7ae52f100613', 'width': 1200}, 'variants': {}}]}
|
Training to Control Test-Time Reasoning Budgets
| 3 |
In reasoning-intensive tasks, optimizing **test-time compute scaling** may offer more leverage than going with the next larger model in the family.
Inspired by Qwen 3's findings, we're exploring how to make reasoning **elastic,** training models that think harder **when requested**.
But simply prompting our finetune for more words isn't enough. Our early dataset (*SpaceThinker*) trained models on short reasoning traces (\~200 tokens), which conditioned them to stop early—even with long contexts and explicit requests for more.
So we're sharing **SpaceOm**: a dataset designed to train **budget-aware reasoning**—models that modulate their thought depth based on prompt constraints like:
\> “Explain briefly” vs. “Give me the full breakdown (\~3000 words)”
This approach taps into the model’s latent capacity to scale reasoning **without scaling model size**—ideal for local deployments in robotics, navigation, and planning, where compute is tight but compositional reasoning is critical.
More details here: [https://remyxai.substack.com/p/use-your-words](https://remyxai.substack.com/p/use-your-words)
| 2025-05-03T21:01:45 |
remyxai
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke32xa
| false | null |
t3_1ke32xa
|
/r/LocalLLaMA/comments/1ke32xa/training_to_control_testtime_reasoning_budgets/
| false | false | 3 |
{'enabled': True, 'images': [{'id': 'RMJa5bLSdNXMiC2QyN9ECYkx-3jvFwNnHafQ5lG4KgI', 'resolutions': [{'height': 67, 'url': 'https://preview.redd.it/jc2e3qt8nmye1.png?width=108&crop=smart&auto=webp&s=b313b5f15e9c49173d9a316cddc34e85ca9ef26a', 'width': 108}, {'height': 135, 'url': 'https://preview.redd.it/jc2e3qt8nmye1.png?width=216&crop=smart&auto=webp&s=753a082ea9d81db5cca1d33778c330bdeac122c9', 'width': 216}, {'height': 201, 'url': 'https://preview.redd.it/jc2e3qt8nmye1.png?width=320&crop=smart&auto=webp&s=8927f3e36038bfdfc70c76610ea7400bf202f698', 'width': 320}, {'height': 402, 'url': 'https://preview.redd.it/jc2e3qt8nmye1.png?width=640&crop=smart&auto=webp&s=741c75ff26137c8f1c34c8b7b860475b6459ce28', 'width': 640}], 'source': {'height': 597, 'url': 'https://preview.redd.it/jc2e3qt8nmye1.png?auto=webp&s=4c1abf19799e4aa6fe170692325815a4c291732f', 'width': 949}, 'variants': {}}]}
|
||
What the actual **** is that? cppscripts.com
| 0 |
So, I wanted to find a lil guide on how to set up llama.cpp to run an LLM locally and to understand what is ollama and what llama.cpp and I found this... which is... something, for sure...
That's what reading about tech without knowing shit feels like, like that "how english sounds to non speakers".
[https://cppscripts.com/llamacpp-vs-ollama](https://cppscripts.com/llamacpp-vs-ollama)
| 2025-05-03T21:06:01 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke36as/what_the_actual_is_that_cppscriptscom/
|
uForgot_urFloaties
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke36as
| false | null |
t3_1ke36as
|
/r/LocalLLaMA/comments/1ke36as/what_the_actual_is_that_cppscriptscom/
| false | false |
self
| 0 |
{'enabled': False, 'images': [{'id': '175mkKaBd8Ywrl5oG-rtvZev_SY3-nB1sXDUHwscy3M', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/z4kmgH__gayAZN9R_jG9VQBSFL6qzcRfn78aHXxaSz8.jpg?width=108&crop=smart&auto=webp&s=1b92c28aa28c1b5ebc4d4a0f2235804d6beec453', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/z4kmgH__gayAZN9R_jG9VQBSFL6qzcRfn78aHXxaSz8.jpg?width=216&crop=smart&auto=webp&s=84f6be251e04b277bca480b92331385b77e2882a', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/z4kmgH__gayAZN9R_jG9VQBSFL6qzcRfn78aHXxaSz8.jpg?width=320&crop=smart&auto=webp&s=1eaa92bfb01a84c4ac10df7d234d69b0b163ff0f', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/z4kmgH__gayAZN9R_jG9VQBSFL6qzcRfn78aHXxaSz8.jpg?width=640&crop=smart&auto=webp&s=450dfb65402d725cc3d83a1c8298fd242f715ff8', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/z4kmgH__gayAZN9R_jG9VQBSFL6qzcRfn78aHXxaSz8.jpg?width=960&crop=smart&auto=webp&s=51f540d3a357b6af102d0af7d5aef7a7419a16ad', 'width': 960}], 'source': {'height': 576, 'url': 'https://external-preview.redd.it/z4kmgH__gayAZN9R_jG9VQBSFL6qzcRfn78aHXxaSz8.jpg?auto=webp&s=01ca52e2371050a989747a5d66aa98383ca3a216', 'width': 1024}, 'variants': {}}]}
|
How is your experience with Qwen3 so far?
| 177 |
Do they prove their worth? Are the benchmark scores appropriate to real world performance?
| 2025-05-03T21:20:12 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke3heg/how_is_your_experience_with_qwen3_so_far/
|
Balance-
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke3heg
| false | null |
t3_1ke3heg
|
/r/LocalLLaMA/comments/1ke3heg/how_is_your_experience_with_qwen3_so_far/
| false | false |
self
| 177 | null |
Qwen 3 235b beats sonnet 3.7 in aider polyglot
| 397 |
Win for open source
| 2025-05-03T21:28:35 |
Independent-Wind4462
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke3nw3
| false | null |
t3_1ke3nw3
|
/r/LocalLLaMA/comments/1ke3nw3/qwen_3_235b_beats_sonnet_37_in_aider_polyglot/
| false | false | 397 |
{'enabled': True, 'images': [{'id': 'zHeE-flQ3RL7XlwuuOWski-SbXBWMC1vnolEtCjJmQs', 'resolutions': [{'height': 119, 'url': 'https://preview.redd.it/sjw1h2yexmye1.jpeg?width=108&crop=smart&auto=webp&s=1a4db4fd2f555e874ba34ff43be41772b503c38b', 'width': 108}, {'height': 239, 'url': 'https://preview.redd.it/sjw1h2yexmye1.jpeg?width=216&crop=smart&auto=webp&s=7e0c9d170d982f4eb49fa69a1272643e943ae0ef', 'width': 216}, {'height': 354, 'url': 'https://preview.redd.it/sjw1h2yexmye1.jpeg?width=320&crop=smart&auto=webp&s=92a72a4e597961660eacbb4bbf7738602b934813', 'width': 320}, {'height': 708, 'url': 'https://preview.redd.it/sjw1h2yexmye1.jpeg?width=640&crop=smart&auto=webp&s=3228950c23440471a53b0440043d2bbfaeaa2b03', 'width': 640}], 'source': {'height': 916, 'url': 'https://preview.redd.it/sjw1h2yexmye1.jpeg?auto=webp&s=573e31ae6fc7adc8ecd9b38f8cc222577bc36bfe', 'width': 827}, 'variants': {}}]}
|
||
Recommend AI’s uncensored
| 1 |
[removed]
| 2025-05-03T21:32:00 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke3qir/recommend_ais_uncensored/
|
Every_Ad2629
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke3qir
| false | null |
t3_1ke3qir
|
/r/LocalLLaMA/comments/1ke3qir/recommend_ais_uncensored/
| false | false |
self
| 1 | null |
How to Prevent Slop and Repetitiveness?
| 1 |
[removed]
| 2025-05-03T21:35:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke3t64/how_to_prevent_slop_and_repetitiveness/
|
Express_Seesaw_8418
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke3t64
| false | null |
t3_1ke3t64
|
/r/LocalLLaMA/comments/1ke3t64/how_to_prevent_slop_and_repetitiveness/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'FJhEfE9TJdtegtJ1mIJD54T4ILqPfGygq9u2oQExrM8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/QTxv3QFdM74MrFyxF-CEW6QhjfswQarknflUW9bjS3o.jpg?width=108&crop=smart&auto=webp&s=193b08caaabe9212083582d4325189bab10687ea', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/QTxv3QFdM74MrFyxF-CEW6QhjfswQarknflUW9bjS3o.jpg?width=216&crop=smart&auto=webp&s=de14a4943c25e741221cb92baf8530f043a12dcb', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/QTxv3QFdM74MrFyxF-CEW6QhjfswQarknflUW9bjS3o.jpg?width=320&crop=smart&auto=webp&s=ef4349fe1ea44f8d2a0156fb21b61a9039d0a97d', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/QTxv3QFdM74MrFyxF-CEW6QhjfswQarknflUW9bjS3o.jpg?width=640&crop=smart&auto=webp&s=34f542e6d4593eebaeecb7ce1becbe5ee89dea61', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/QTxv3QFdM74MrFyxF-CEW6QhjfswQarknflUW9bjS3o.jpg?width=960&crop=smart&auto=webp&s=ee38f057b95629c0c107e127986b77d1f811b92c', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/QTxv3QFdM74MrFyxF-CEW6QhjfswQarknflUW9bjS3o.jpg?width=1080&crop=smart&auto=webp&s=edd11230e1970d21d6a9786a50c053f86e0f6600', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/QTxv3QFdM74MrFyxF-CEW6QhjfswQarknflUW9bjS3o.jpg?auto=webp&s=934569687a46f4a7758d9aa0d6e5aebae23e4cf1', 'width': 1200}, 'variants': {}}]}
|
Best model for coding that can run on MacBook M1 Max 64GB unified memory, with 32 GPU cores
| 1 |
[removed]
| 2025-05-03T21:50:52 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke453d/best_model_for_coding_that_can_run_on_macbook_m1/
|
tghrowawayg
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke453d
| false | null |
t3_1ke453d
|
/r/LocalLLaMA/comments/1ke453d/best_model_for_coding_that_can_run_on_macbook_m1/
| false | false |
self
| 1 | null |
Train model on telegram messages
| 1 |
[removed]
| 2025-05-03T21:55:05 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke487c/train_model_on_telegram_messages/
|
Evening-Anteater-519
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke487c
| false | null |
t3_1ke487c
|
/r/LocalLLaMA/comments/1ke487c/train_model_on_telegram_messages/
| false | false |
self
| 1 | null |
https://www.freelance.wtf/
| 1 |
[removed]
| 2025-05-03T21:55:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke48cl/httpswwwfreelancewtf/
|
OkClerk7966
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke48cl
| false | null |
t3_1ke48cl
|
/r/LocalLLaMA/comments/1ke48cl/httpswwwfreelancewtf/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'wELAhY2qRF4StUfAEpkujPg3jJJhHxjy9ORHQK-QiV8', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/AjwOZJ9DemF2UBLkfd95vy7b3lYZyuSIAFaDXWXP_gI.jpg?width=108&crop=smart&auto=webp&s=02327cbd9c298f22ba972c832c298496c72bb38d', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/AjwOZJ9DemF2UBLkfd95vy7b3lYZyuSIAFaDXWXP_gI.jpg?width=216&crop=smart&auto=webp&s=3e0d1638dc4a4d1cefac34a9b72bfca157578600', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/AjwOZJ9DemF2UBLkfd95vy7b3lYZyuSIAFaDXWXP_gI.jpg?width=320&crop=smart&auto=webp&s=b8933629c4ef70a188dde5fab3ee85f7bd8e969a', 'width': 320}], 'source': {'height': 500, 'url': 'https://external-preview.redd.it/AjwOZJ9DemF2UBLkfd95vy7b3lYZyuSIAFaDXWXP_gI.jpg?auto=webp&s=7e600e6655ea713cb4e52677a4699e37b9e93197', 'width': 500}, 'variants': {}}]}
|
What's the best benchmarking website right now in your opinion?
| 1 |
[removed]
| 2025-05-03T21:57:50 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke4a8o/whats_the_best_benchmarking_website_right_now_in/
|
UnstoppableGooner
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke4a8o
| false | null |
t3_1ke4a8o
|
/r/LocalLLaMA/comments/1ke4a8o/whats_the_best_benchmarking_website_right_now_in/
| false | false |
self
| 1 | null |
Personal fav tool for generation
| 1 |
[removed]
| 2025-05-03T22:03:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke4eog/personal_fav_tool_for_generation/
|
Apart_Click_6938
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke4eog
| false | null |
t3_1ke4eog
|
/r/LocalLLaMA/comments/1ke4eog/personal_fav_tool_for_generation/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'VSiqbcc5ID-xayz4c5tFj4-RodarvgiQazOvcgeZzxY', 'resolutions': [{'height': 56, 'url': 'https://external-preview.redd.it/RAJzRUqG9qsWVxgSsqlRoVa3ozD6HQ2QYkjQGx_fTfU.jpg?width=108&crop=smart&auto=webp&s=ae9d24eafca4c61499fcdba5935fe07f1655fac3', 'width': 108}, {'height': 113, 'url': 'https://external-preview.redd.it/RAJzRUqG9qsWVxgSsqlRoVa3ozD6HQ2QYkjQGx_fTfU.jpg?width=216&crop=smart&auto=webp&s=ef03414335a948ef870a16b76cb5924c07b1c401', 'width': 216}, {'height': 168, 'url': 'https://external-preview.redd.it/RAJzRUqG9qsWVxgSsqlRoVa3ozD6HQ2QYkjQGx_fTfU.jpg?width=320&crop=smart&auto=webp&s=2a7bbcdddd3709759c35c79698c7e6bdd4a345c6', 'width': 320}, {'height': 336, 'url': 'https://external-preview.redd.it/RAJzRUqG9qsWVxgSsqlRoVa3ozD6HQ2QYkjQGx_fTfU.jpg?width=640&crop=smart&auto=webp&s=a983fe0a65259dbbb76130d8751532754ad9d0b3', 'width': 640}, {'height': 504, 'url': 'https://external-preview.redd.it/RAJzRUqG9qsWVxgSsqlRoVa3ozD6HQ2QYkjQGx_fTfU.jpg?width=960&crop=smart&auto=webp&s=34a8dd55022b6e1f2cbc6fc2642d87c17a900ace', 'width': 960}, {'height': 567, 'url': 'https://external-preview.redd.it/RAJzRUqG9qsWVxgSsqlRoVa3ozD6HQ2QYkjQGx_fTfU.jpg?width=1080&crop=smart&auto=webp&s=22b460b1a487b64db8980f6d4df5ad5ba12aece6', 'width': 1080}], 'source': {'height': 630, 'url': 'https://external-preview.redd.it/RAJzRUqG9qsWVxgSsqlRoVa3ozD6HQ2QYkjQGx_fTfU.jpg?auto=webp&s=7efca27fe98e854362eea866899e912d42649930', 'width': 1200}, 'variants': {}}]}
|
How to get around slow prompt eval?
| 5 |
I'm running Qwen2.5 Coder 1.5B on my Ryzen 5 5625U APU using llama.cpp and Vulkan. I would like to use it as a code completion modal, however, I only get about 30t/s on prompt evaluation.
This means that ingesting a whole code file and generating a completion takes a lot of time, especially as context fills up.
I've tried the Continue.dev and llama.vscode extensions. The latter is more lightweight, but doesn't cancel the previous request when the file is modified.
Is there a way I can make local models more usable for code autocomplete? Should I perhaps try another engine? Is a newer MoE model going to have faster PP?
| 2025-05-03T22:10:31 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke4juq/how_to_get_around_slow_prompt_eval/
|
antonlyap
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke4juq
| false | null |
t3_1ke4juq
|
/r/LocalLLaMA/comments/1ke4juq/how_to_get_around_slow_prompt_eval/
| false | false |
self
| 5 | null |
It would be nice to have a wiki on this sub.
| 1 |
[removed]
| 2025-05-03T22:52:55 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke5ex2/it_would_be_nice_to_have_a_wiki_on_this_sub/
|
Kyla_3049
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke5ex2
| false | null |
t3_1ke5ex2
|
/r/LocalLLaMA/comments/1ke5ex2/it_would_be_nice_to_have_a_wiki_on_this_sub/
| false | false |
self
| 1 | null |
The More I Use AI, the More I Worry About My Privacy
| 1 |
[removed]
| 2025-05-03T23:01:04 |
[deleted]
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke5kwi
| false | null |
t3_1ke5kwi
|
/r/LocalLLaMA/comments/1ke5kwi/the_more_i_use_ai_the_more_i_worry_about_my/
| false | false |
default
| 1 | null |
||
[M3 Ultra 512GB] LM Studio + GGUF + Qwen3-235B-A22B_Q8+reasoning
| 6 |
I just run the setup mentioned in tittle. I have questioned it about the Characterization of Carbon Nanotubes as I have worked in two old publications about it and asked it to answer as a PhD.
Well, it runned the prompt at 15 tokens/sec. what impressed me, in a mixed way, is that it recognized old chats I had, even if I was in a new chat and in a new folder, also it only uses 3-12% of the CPU.
The response and "thinking" was highly coherent.
| 2025-05-03T23:50:53 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke6l65/m3_ultra_512gb_lm_studio_gguf_qwen3235ba22b/
|
Turbulent_Pin7635
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke6l65
| false | null |
t3_1ke6l65
|
/r/LocalLLaMA/comments/1ke6l65/m3_ultra_512gb_lm_studio_gguf_qwen3235ba22b/
| false | false |
self
| 6 | null |
New LLM with no training - Minimum Viable Product
| 0 |
Someone I know created this LLM without training at all
[www.dopegpt.com](http://www.dopegpt.com)
It is a mimium viable product that can easily be improved upon. Comment below your thoughts
| 2025-05-04T00:05:15 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke6vh4/new_llm_with_no_training_minimum_viable_product/
|
6whiten_igga9
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke6vh4
| false | null |
t3_1ke6vh4
|
/r/LocalLLaMA/comments/1ke6vh4/new_llm_with_no_training_minimum_viable_product/
| false | false |
self
| 0 | null |
What happened after original ChatGPT that models started improving exponentially?
| 35 |
It seems like till GPT3.5 and ChatGPT model development was rather slow and a niche field of computer science.
Suddenly after that model development has supercharged.
Were big tech companies just sitting on this capability, but not building because they thought it would be too expensive and couldn't figure a product strategy around this?
| 2025-05-04T00:27:41 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke7aqj/what_happened_after_original_chatgpt_that_models/
|
Euphoric_Sandwich_74
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke7aqj
| false | null |
t3_1ke7aqj
|
/r/LocalLLaMA/comments/1ke7aqj/what_happened_after_original_chatgpt_that_models/
| false | false |
self
| 35 | null |
llama.cpp now supports Llama-3_1-Nemotron-Ultra-253B-v1
| 65 |
llama.cpp now supports Nvidia's Llama-3\_1-Nemotron-Ultra-253B-v1 starting from b5270.
[https://github.com/ggml-org/llama.cpp/pull/12843](https://github.com/ggml-org/llama.cpp/pull/12843)
Supposedly it is better than DeepSeek R1:
[https://www.reddit.com/r/LocalLLaMA/comments/1ju6sm1/nvidiallama3\_1nemotronultra253bv1\_hugging\_face/](https://www.reddit.com/r/LocalLLaMA/comments/1ju6sm1/nvidiallama3_1nemotronultra253bv1_hugging_face/)
It is the biggest SOTA dense model with reasoning fine tune now. So it is worth it to explore what it does best comparing to other models.
Model size is 38% smaller than the source Llama-3.1-405B. KV cache is 49% smaller. Overall, memory footprint is 39% smaller at 128k context.
IQ3\_M should be around 110GB. While fp16 KV cache is 32GB at 128k, IQ4\_NL KV cahce is only 9GB at 128k context. Seems like a perfect fit for >=128GB Apple Silicon or the upcoming DGX Spark.
If you have the resource to run this model, give it a try and see if it can beat DeepSeek R1 as they claim!
PS Nemotron pruned models in general are good when you can load it fully to your VRAM. However, it suffers from uneven VRAM distribution when you have multiple cards. To get around that, it is recommended that you tinker with the "-ts" switch to set VRAM distribution manually until someone implemented automatic VRAM distribution.
[https://github.com/ggml-org/llama.cpp/issues/12654](https://github.com/ggml-org/llama.cpp/issues/12654)
I made an Excel to breakdown the exact amount of VRAM usage for each layer. It can serve as a starting point for you to set "-ts" if you have multiple cards.
[https://huggingface.co/ymcki/Llama-3\_1-Nemotron-51B-Instruct-GGUF/resolve/main/deci.xlsx?download=true](https://huggingface.co/ymcki/Llama-3_1-Nemotron-51B-Instruct-GGUF/resolve/main/deci.xlsx?download=true)
| 2025-05-04T00:35:07 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke7fli/llamacpp_now_supports_llama3_1nemotronultra253bv1/
|
Ok_Warning2146
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke7fli
| false | null |
t3_1ke7fli
|
/r/LocalLLaMA/comments/1ke7fli/llamacpp_now_supports_llama3_1nemotronultra253bv1/
| false | false |
self
| 65 |
{'enabled': False, 'images': [{'id': 'VjTnTmYiVaALHZPmS5-0-GlJr2g0Zrlve2xASwMUHY8', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/dY0oDWmssevMQHDIEOS75Izth4TEVBBqynZPVGzCma8.jpg?width=108&crop=smart&auto=webp&s=cf84baf857057982711471a5046a23f5a84474b6', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/dY0oDWmssevMQHDIEOS75Izth4TEVBBqynZPVGzCma8.jpg?width=216&crop=smart&auto=webp&s=ba7bbf397492e5f71c6b346724f7dec369597798', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/dY0oDWmssevMQHDIEOS75Izth4TEVBBqynZPVGzCma8.jpg?width=320&crop=smart&auto=webp&s=06cb78a892b8821adfcbcb69d95dc9ac3a49e9ff', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/dY0oDWmssevMQHDIEOS75Izth4TEVBBqynZPVGzCma8.jpg?width=640&crop=smart&auto=webp&s=8c62ae4e9383b9a9cf52ed7606da98d99da05586', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/dY0oDWmssevMQHDIEOS75Izth4TEVBBqynZPVGzCma8.jpg?width=960&crop=smart&auto=webp&s=a90f8e43e75de4439bbce51ecd406d4633655713', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/dY0oDWmssevMQHDIEOS75Izth4TEVBBqynZPVGzCma8.jpg?width=1080&crop=smart&auto=webp&s=b66dd3c65cb16b8a10d8d7304adcb6c6ca3314a4', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/dY0oDWmssevMQHDIEOS75Izth4TEVBBqynZPVGzCma8.jpg?auto=webp&s=b4e2a1632c23a92fd12970ee5a551b2c93892dc8', 'width': 1200}, 'variants': {}}]}
|
The grok2 system prompt actually works pretty well with Qwen 2.5 (atleast on 14b)
| 0 |
I did have to modify it to fit the capabilities the frontend I use has, however.
| 2025-05-04T00:43:11 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke7ku3/the_grok2_system_prompt_actually_works_pretty/
|
__laughing__
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke7ku3
| false | null |
t3_1ke7ku3
|
/r/LocalLLaMA/comments/1ke7ku3/the_grok2_system_prompt_actually_works_pretty/
| false | false |
self
| 0 | null |
Aider - Qwen 32b 64% !
| 1 |
[deleted]
| 2025-05-04T00:54:12 |
[deleted]
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke7rx9
| false | null |
t3_1ke7rx9
|
/r/LocalLLaMA/comments/1ke7rx9/aider_qwen_32b_64/
| false | false |
default
| 1 | null |
||
Aider - qwen 32b 45% !
| 75 |
link
[Add benchmarks for Qwen3-235B-A22B and Qwen3-32B by AlongWY · Pull Request #3908 · Aider-AI/aider · GitHub](https://github.com/Aider-AI/aider/pull/3908/files)
| 2025-05-04T00:55:37 |
Healthy-Nebula-3603
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke7ssw
| false | null |
t3_1ke7ssw
|
/r/LocalLLaMA/comments/1ke7ssw/aider_qwen_32b_45/
| false | false | 75 |
{'enabled': True, 'images': [{'id': '5B1dbbXCV8oVvn8p6LEyYW-I2oNrHSw-uv1SR9R_E7o', 'resolutions': [{'height': 139, 'url': 'https://preview.redd.it/9k3fxxt9ynye1.png?width=108&crop=smart&auto=webp&s=a60c54302b70f561447a550d1c40172de9dc7d23', 'width': 108}, {'height': 279, 'url': 'https://preview.redd.it/9k3fxxt9ynye1.png?width=216&crop=smart&auto=webp&s=69f11db2df530b81cd141251491c709ab2e2ed35', 'width': 216}, {'height': 414, 'url': 'https://preview.redd.it/9k3fxxt9ynye1.png?width=320&crop=smart&auto=webp&s=91d04af4b4905e1fe5ea7400c8fbfcc3c64cb513', 'width': 320}, {'height': 828, 'url': 'https://preview.redd.it/9k3fxxt9ynye1.png?width=640&crop=smart&auto=webp&s=1e9ca848eede216b3f41bc1736fb709414c67f6f', 'width': 640}, {'height': 1243, 'url': 'https://preview.redd.it/9k3fxxt9ynye1.png?width=960&crop=smart&auto=webp&s=3836e6763347621386a84ea2fbc9efe9a4081311', 'width': 960}], 'source': {'height': 1246, 'url': 'https://preview.redd.it/9k3fxxt9ynye1.png?auto=webp&s=cf2e6ffa4442e79c8d0058e042c9f90611d1a85f', 'width': 962}, 'variants': {}}]}
|
||
What’s your favorite GUI
| 43 |
Can be web based or app like LM Studio
Can be local llm only or able to connect online api like openai, openrouter, etc
Trying to learn about new tools
| 2025-05-04T01:08:42 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke81ai/whats_your_favorite_gui/
|
Dentifrice
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke81ai
| false | null |
t3_1ke81ai
|
/r/LocalLLaMA/comments/1ke81ai/whats_your_favorite_gui/
| false | false |
self
| 43 | null |
Any in-depth tutorials which do step-by-step walkthroughs on how to fine-tune an LLM?
| 43 |
Hi!
I want to learn about the full process, from soup to nuts, of how to fine-tune an LLM. If anyone has well-documented resources, videos, or tutorials that they could point me to, that would be spectacular.
If there are also related resources about LLMs' benchmarking and evaluations, that would be incredibly helpful as well.
Thank you!!
| 2025-05-04T01:10:40 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke82nc/any_indepth_tutorials_which_do_stepbystep/
|
darkGrayAdventurer
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke82nc
| false | null |
t3_1ke82nc
|
/r/LocalLLaMA/comments/1ke82nc/any_indepth_tutorials_which_do_stepbystep/
| false | false |
self
| 43 | null |
Keep dancing #beauti #lifeisbutadream #whowillbemylifepartnerta #fashion...
| 1 |
[removed]
| 2025-05-04T01:43:30 |
https://youtube.com/watch?v=tO5m5mwtElw&si=ww_WXLITBsfZrxBN
|
Ok-Maize-4629
|
youtube.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke8nn9
| false |
{'oembed': {'author_name': 'Ada', 'author_url': 'https://www.youtube.com/@Ada-y5f6o', 'height': 200, 'html': '<iframe width="356" height="200" src="https://www.youtube.com/embed/tO5m5mwtElw?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="Keep dancing #beauti #lifeisbutadream #whowillbemylifepartnerta #fashion #dj #dance"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/tO5m5mwtElw/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'Keep dancing #beauti #lifeisbutadream #whowillbemylifepartnerta #fashion #dj #dance', 'type': 'video', 'version': '1.0', 'width': 356}, 'type': 'youtube.com'}
|
t3_1ke8nn9
|
/r/LocalLLaMA/comments/1ke8nn9/keep_dancing_beauti_lifeisbutadream/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'bmx_3UX3eZFq6ewrnCgIrpltglkTC_zZa37qy_He7WI', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/y22Pg6ViCcmAbnbK5NVBk_zCefKTHgVYjo0ie21mOFY.jpg?width=108&crop=smart&auto=webp&s=04ce95125360c1f12c0bda57b696901f19eea437', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/y22Pg6ViCcmAbnbK5NVBk_zCefKTHgVYjo0ie21mOFY.jpg?width=216&crop=smart&auto=webp&s=b817c5c06037fb71a76ad1e11b2a22f9ba7cc9d0', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/y22Pg6ViCcmAbnbK5NVBk_zCefKTHgVYjo0ie21mOFY.jpg?width=320&crop=smart&auto=webp&s=66f58080969a7de9a3b308137e6d6db16951c13a', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/y22Pg6ViCcmAbnbK5NVBk_zCefKTHgVYjo0ie21mOFY.jpg?auto=webp&s=8ca1895177f2f48de60c06a609a6e719326de428', 'width': 480}, 'variants': {}}]}
|
|
Help with AI video generation
| 1 |
[removed]
| 2025-05-04T02:05:21 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke91qd/help_with_ai_video_generation/
|
MoreCat4994
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke91qd
| false | null |
t3_1ke91qd
|
/r/LocalLLaMA/comments/1ke91qd/help_with_ai_video_generation/
| false | false |
self
| 1 | null |
What's the most reliable/best benchmarking website currently, in your opinion?
| 1 |
[removed]
| 2025-05-04T02:21:50 |
[deleted]
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke9caf
| false | null |
t3_1ke9caf
|
/r/LocalLLaMA/comments/1ke9caf/whats_the_most_reliablebest_benchmarking_website/
| false | false |
default
| 1 | null |
||
My project
| 1 |
[removed]
| 2025-05-04T02:23:41 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke9dgv/my_project/
|
Previous-Amphibian23
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke9dgv
| false | null |
t3_1ke9dgv
|
/r/LocalLLaMA/comments/1ke9dgv/my_project/
| false | false |
self
| 1 | null |
UI-Tars-1.5 reasoning never fails to entertain me.
| 1 |
[removed]
| 2025-05-04T02:33:46 |
Successful_Bowl2564
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke9jmo
| false | null |
t3_1ke9jmo
|
/r/LocalLLaMA/comments/1ke9jmo/uitars15_reasoning_never_fails_to_entertain_me/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'n0VJIBD7SMNgIPSjeUyHXUF8Hhz6aa8V8SIwlGAvAxo', 'resolutions': [{'height': 69, 'url': 'https://preview.redd.it/v28k9u0rfoye1.png?width=108&crop=smart&auto=webp&s=f49d1f252e82b752466ef84e5ed5d305fca2a7b5', 'width': 108}, {'height': 138, 'url': 'https://preview.redd.it/v28k9u0rfoye1.png?width=216&crop=smart&auto=webp&s=d839aaefe15059817e324d25c79f12c9454b39b8', 'width': 216}, {'height': 204, 'url': 'https://preview.redd.it/v28k9u0rfoye1.png?width=320&crop=smart&auto=webp&s=20cd3f625d63e9f15fdbb6f1951d824328717a77', 'width': 320}, {'height': 409, 'url': 'https://preview.redd.it/v28k9u0rfoye1.png?width=640&crop=smart&auto=webp&s=d2cdbca440b4aaa7e259168d61ac3e1e1aa1796e', 'width': 640}], 'source': {'height': 466, 'url': 'https://preview.redd.it/v28k9u0rfoye1.png?auto=webp&s=a1b6568f9d2e11368d8d7d731ec698cd5b54b126', 'width': 729}, 'variants': {}}]}
|
||
MCP_A2A - Use A2A Agents from MCP Clients
| 1 |
I couldn't find any , so I created this quick and dirty MCP Server to allow me to communicate with A2A agents, sharing here just in case someone finds it useful.
It is definitely not production ready, and I will improve it a little bit more for my personal needs... but it might help someone who is starting a project like I was today... If something is wrong it is very easy to change as the lib is very simple , hopefully works out-of-the-box probably out-of-the-spec. Enjoy.
| 2025-05-04T02:48:06 |
https://github.com/regismesquita/MCP_A2A
|
das_rdsm
|
github.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke9sqy
| false | null |
t3_1ke9sqy
|
/r/LocalLLaMA/comments/1ke9sqy/mcp_a2a_use_a2a_agents_from_mcp_clients/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'EgjPfU9ampPrhXR5IpD0WMP_RTsH5kUFitpXyyjHPUU', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/ctqhy8LiC1YrNpDOh5mnILsLKlqVmzRID6WwJPWaepg.jpg?width=108&crop=smart&auto=webp&s=9e9aaf8a7ccfca06c5564b3f018005f45ed79c03', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/ctqhy8LiC1YrNpDOh5mnILsLKlqVmzRID6WwJPWaepg.jpg?width=216&crop=smart&auto=webp&s=a787f04f10236f5a46ff00255a3852c178205957', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/ctqhy8LiC1YrNpDOh5mnILsLKlqVmzRID6WwJPWaepg.jpg?width=320&crop=smart&auto=webp&s=cb37d6a6eb0ce94b7503eba3c156318c9586ba0f', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/ctqhy8LiC1YrNpDOh5mnILsLKlqVmzRID6WwJPWaepg.jpg?width=640&crop=smart&auto=webp&s=206d6de7cc80c381197c33612fadac856b1fcef4', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/ctqhy8LiC1YrNpDOh5mnILsLKlqVmzRID6WwJPWaepg.jpg?width=960&crop=smart&auto=webp&s=ef15dba3e05574ea490f6b0d6023ccef9a0c273e', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/ctqhy8LiC1YrNpDOh5mnILsLKlqVmzRID6WwJPWaepg.jpg?width=1080&crop=smart&auto=webp&s=c2d80b2c74eb1e8f24312c1123f5011c298d420e', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/ctqhy8LiC1YrNpDOh5mnILsLKlqVmzRID6WwJPWaepg.jpg?auto=webp&s=597be702a8b44f6952b012bb1897b9bb796022e4', 'width': 1200}, 'variants': {}}]}
|
|
A question for fellow 48gb RTX 4090D owners
| 6 |
I have the chinese blower 48gb rtx 4090D and the vBIOS has it locked to prevent fan from going under 30%, and by default it won't idle the memory clock and keeps it at 10,500mhz which wastes a lot of power.
The memory clock can be fixed by manually setting it down to 405mhz which helps the idle power usage, but not so much the noise from the fan always at 30%. Disabling the gpu in device manager does make the fan idle very quietly but then the power usage jumps up by about 50W again.
Any ways to update the vBIOS to fix these slight gripes?
| 2025-05-04T02:54:07 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke9wgs/a_question_for_fellow_48gb_rtx_4090d_owners/
|
Goldkoron
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke9wgs
| false | null |
t3_1ke9wgs
|
/r/LocalLLaMA/comments/1ke9wgs/a_question_for_fellow_48gb_rtx_4090d_owners/
| false | false |
self
| 6 | null |
Should i get an RX 7900 XTX as a Linux gamer that also enjoys using local AIs?
| 2 |
Hey so i'm upgrading from an RTX 3070.
I'm torn between a used RTX 3090 and used RX 7900 XTX as an upgrade. TDP and price are the same.
Of course i know that for pure AI work that Nvidia is still the king due to CUDA being everywhere.
But on the other hand, AMD offers quite a bit more performance in games.
So my question is if the RX 7900 XTX card would be good enough to get decent token speeds for quantized 24B/32B models and at least as fast image generation as my RTX 3070. Like essentially how hard is it for the end user to work with ROCm on Linux nowadays?
For reference i'm currently using LMStudio and StabilityMatrix for my AI needs.
Similarly, would i be able to combine my two cards so part of the model runs on CUDA and the other on ROCM? Would this help?
| 2025-05-04T02:57:05 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke9ybp/should_i_get_an_rx_7900_xtx_as_a_linux_gamer_that/
|
HRudy94
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke9ybp
| false | null |
t3_1ke9ybp
|
/r/LocalLLaMA/comments/1ke9ybp/should_i_get_an_rx_7900_xtx_as_a_linux_gamer_that/
| false | false |
self
| 2 | null |
2nd Brain
| 0 |
What's the consensus in building a 2nd brain? I think everyone will eventually have their own agent that will do things for you in the digital space with all your data stored locally, but I'm wondering what the best practices/techniques are regarding this?
Right now I'm deciding between \[Obsidian as a store of data and using RAG and some techniques to make the output better \] or building my own system from the ground up.
I don't have anything completed fully yet, but I did play around with NotebookLM and I realized it does a really good job with generating that conversational podcast as well as having their own built in functions like mindmaps, you can query the documents, etc.
Overall I want to move forward with building this second brain, such as the ingestion, chunking, embedding, and rag part but appreciate opinions on what others are thinking and doing, thanks!
| 2025-05-04T02:57:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1ke9yh5/2nd_brain/
|
proxyplz
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ke9yh5
| false | null |
t3_1ke9yh5
|
/r/LocalLLaMA/comments/1ke9yh5/2nd_brain/
| false | false |
self
| 0 | null |
UI-Tars-1.5 reasoning never fails to entertain me
| 1 |
[removed]
| 2025-05-04T03:21:03 |
Successful_Bowl2564
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kead2j
| false | null |
t3_1kead2j
|
/r/LocalLLaMA/comments/1kead2j/uitars15_reasoning_never_fails_to_entertain_me/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'vl-5Cyhwgv15C91XIykPbrMhdcRCu_AhZ0brPpuHA6M', 'resolutions': [{'height': 69, 'url': 'https://preview.redd.it/q4nvjdi6ooye1.png?width=108&crop=smart&auto=webp&s=da0bae45c0eb69734951c0cd9bdb07b4209b6c88', 'width': 108}, {'height': 138, 'url': 'https://preview.redd.it/q4nvjdi6ooye1.png?width=216&crop=smart&auto=webp&s=5da634bf6da088c09e52a7fe74e641796a61c496', 'width': 216}, {'height': 204, 'url': 'https://preview.redd.it/q4nvjdi6ooye1.png?width=320&crop=smart&auto=webp&s=133a01724a52fdd68652611eaf603373b8d60a9f', 'width': 320}, {'height': 409, 'url': 'https://preview.redd.it/q4nvjdi6ooye1.png?width=640&crop=smart&auto=webp&s=618fa306dd9347e7139019d08fcdc5ffd5583df0', 'width': 640}], 'source': {'height': 466, 'url': 'https://preview.redd.it/q4nvjdi6ooye1.png?auto=webp&s=352f8bcd0c370bcaea58b5840a6182d7eac2d930', 'width': 729}, 'variants': {}}]}
|
||
UI-Tars-1.5 reasoning never fails to entertain me
| 1 |
[removed]
| 2025-05-04T03:28:46 |
Successful_Bowl2564
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1keaho3
| false | null |
t3_1keaho3
|
/r/LocalLLaMA/comments/1keaho3/uitars15_reasoning_never_fails_to_entertain_me/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'vl-5Cyhwgv15C91XIykPbrMhdcRCu_AhZ0brPpuHA6M', 'resolutions': [{'height': 69, 'url': 'https://preview.redd.it/w2e9jb4mpoye1.png?width=108&crop=smart&auto=webp&s=337302a4a6991ccccb6a801ea987fe3420d6beae', 'width': 108}, {'height': 138, 'url': 'https://preview.redd.it/w2e9jb4mpoye1.png?width=216&crop=smart&auto=webp&s=162949a113a1a5aa3a3deabadb4c9cbd23330eca', 'width': 216}, {'height': 204, 'url': 'https://preview.redd.it/w2e9jb4mpoye1.png?width=320&crop=smart&auto=webp&s=4bae9e4e7c3820e46ecf044513f3067ef5746dee', 'width': 320}, {'height': 409, 'url': 'https://preview.redd.it/w2e9jb4mpoye1.png?width=640&crop=smart&auto=webp&s=cfc7fd1d4bcf37b3e6938bf90b8707d54bd98cd7', 'width': 640}], 'source': {'height': 466, 'url': 'https://preview.redd.it/w2e9jb4mpoye1.png?auto=webp&s=0bfa2f5e772c389243b2b877a37592cc7392a42f', 'width': 729}, 'variants': {}}]}
|
||
Cline or Roo code LLMs with Less than 6 Gb VRAM? 16GB RAM?
| 0 |
Hey guys. I got a rtx 3060 mobile on my laptop. With 16gigs of RAM.
Is there any way I could get 32 k context window in coding with Roo code or cline?
I downloaded a gguf for ollama, a qwen 7b optimised for tool calling and agentic coding. I accidentally made the context 120k lol. It was working fine in the sense it could see files, etc. But it couldn't make any changes. And if I made the context length anything lower than 8k, the tool would be useless in my vast codebase. Any sweet spot that I can hit?
| 2025-05-04T03:29:03 |
https://www.reddit.com/r/LocalLLaMA/comments/1keahu8/cline_or_roo_code_llms_with_less_than_6_gb_vram/
|
backinthe90siwasinav
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1keahu8
| false | null |
t3_1keahu8
|
/r/LocalLLaMA/comments/1keahu8/cline_or_roo_code_llms_with_less_than_6_gb_vram/
| false | false |
self
| 0 | null |
UI-Tars-1.5 reasoning never fails to entertain me
| 1 |
[removed]
| 2025-05-04T03:32:28 |
[deleted]
| 1970-01-01T00:00:00 | 0 |
{}
|
1keajz0
| false | null |
t3_1keajz0
|
/r/LocalLLaMA/comments/1keajz0/uitars15_reasoning_never_fails_to_entertain_me/
| false | false |
default
| 1 | null |
||
Which hardware should I choose for this requirement?
| 2 |
Target performance: 2000t/s Prefill, 100 token/s generation for each user. 10 simultaneous users each with ~50k working context.
Target Model: Qwen3-235B-A22B-Q8_0 at 128k context q8 KV cache.
What is the minimum/cheapest hardware for this requirement on cloud.
| 2025-05-04T03:36:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1keam82/which_hardware_should_i_choose_for_this/
|
Yes_but_I_think
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1keam82
| false | null |
t3_1keam82
|
/r/LocalLLaMA/comments/1keam82/which_hardware_should_i_choose_for_this/
| false | false |
self
| 2 | null |
Apparently shipping AI platforms is a thing now as per this post from the Qwen X account
| 414 | 2025-05-04T04:16:37 |
MushroomGecko
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kebauw
| false | null |
t3_1kebauw
|
/r/LocalLLaMA/comments/1kebauw/apparently_shipping_ai_platforms_is_a_thing_now/
| false | false | 414 |
{'enabled': True, 'images': [{'id': 'F1EczZMgFv8DUSba9m8BfbTe7aRCOYmybdYsDlzzELQ', 'resolutions': [{'height': 43, 'url': 'https://preview.redd.it/fjze9by1yoye1.jpeg?width=108&crop=smart&auto=webp&s=f9b0affac78c522180129a5db84a1496ff7a0554', 'width': 108}, {'height': 86, 'url': 'https://preview.redd.it/fjze9by1yoye1.jpeg?width=216&crop=smart&auto=webp&s=19d9c53df63b7151701fe8873c587b0d63734269', 'width': 216}, {'height': 127, 'url': 'https://preview.redd.it/fjze9by1yoye1.jpeg?width=320&crop=smart&auto=webp&s=808b7c2fb53e37dd27f21807831d5734e3bd1fe8', 'width': 320}, {'height': 255, 'url': 'https://preview.redd.it/fjze9by1yoye1.jpeg?width=640&crop=smart&auto=webp&s=7e0eb0e11cb1eca743ebbe8cd75e69b96fffd960', 'width': 640}, {'height': 382, 'url': 'https://preview.redd.it/fjze9by1yoye1.jpeg?width=960&crop=smart&auto=webp&s=4600b4deb1f5eb625f99fa179e5ca577b95e2600', 'width': 960}, {'height': 430, 'url': 'https://preview.redd.it/fjze9by1yoye1.jpeg?width=1080&crop=smart&auto=webp&s=4ab280c6a88ba7855b053931e27ccf40ef98f02b', 'width': 1080}], 'source': {'height': 816, 'url': 'https://preview.redd.it/fjze9by1yoye1.jpeg?auto=webp&s=b7d96190a5ed7ae43b3d2028516e58076f586047', 'width': 2048}, 'variants': {}}]}
|
|||
next SOTA in vision will be open weights model? when Qwen3 VL?
| 33 |
[https://rank.opencompass.org.cn/leaderboard-multimodal-official/?m=REALTIME](https://rank.opencompass.org.cn/leaderboard-multimodal-official/?m=REALTIME)
| 2025-05-04T04:17:07 |
secopsml
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kebb5e
| false | null |
t3_1kebb5e
|
/r/LocalLLaMA/comments/1kebb5e/next_sota_in_vision_will_be_open_weights_model/
| false | false | 33 |
{'enabled': True, 'images': [{'id': 'O8VIHX-ca0UW00ctbRPU1zhp0hT9QnTQSSdAfk-sUwE', 'resolutions': [{'height': 78, 'url': 'https://preview.redd.it/kshkghcyxoye1.png?width=108&crop=smart&auto=webp&s=40c680bee74f00a331884897580afbe0ce151106', 'width': 108}, {'height': 156, 'url': 'https://preview.redd.it/kshkghcyxoye1.png?width=216&crop=smart&auto=webp&s=3996a58f62114d912359bc2320cee71b8a5c056f', 'width': 216}, {'height': 231, 'url': 'https://preview.redd.it/kshkghcyxoye1.png?width=320&crop=smart&auto=webp&s=765ce9a9873e451d17b2e7cd40eebfeb309ec62a', 'width': 320}, {'height': 462, 'url': 'https://preview.redd.it/kshkghcyxoye1.png?width=640&crop=smart&auto=webp&s=c66c4eab01ae508d363c06770d2b343749af3e76', 'width': 640}, {'height': 693, 'url': 'https://preview.redd.it/kshkghcyxoye1.png?width=960&crop=smart&auto=webp&s=9dc561931c5a433730097d5fa1437aef4e8b55b6', 'width': 960}, {'height': 780, 'url': 'https://preview.redd.it/kshkghcyxoye1.png?width=1080&crop=smart&auto=webp&s=b709cedbdf76beb5123c2b517ba7479ed728aeae', 'width': 1080}], 'source': {'height': 964, 'url': 'https://preview.redd.it/kshkghcyxoye1.png?auto=webp&s=2b92236070b0e3968c5a07fd661a8641d53c89f9', 'width': 1334}, 'variants': {}}]}
|
||
Quick shout-out to Qwen3-30b-a3b as a study tool for Calc2/3
| 90 |
Hi all,
I know the recent Qwen launch has been glazed to death already, but I want to give extra praise and acclaim to this model when it comes to studying. Extremely fast responses of broad, complex topics which are otherwise explained by AWFUL lecturers with terrible speaking skills. Yes, it isnt as smart as the 32b alternative, but for explanations of concepts or integrations/derivations, it is more than enough AND 3x the speed.
Thank you Alibaba,
EEE student.
| 2025-05-04T04:35:32 |
https://www.reddit.com/r/LocalLLaMA/comments/1keblo7/quick_shoutout_to_qwen330ba3b_as_a_study_tool_for/
|
Skkeep
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1keblo7
| false | null |
t3_1keblo7
|
/r/LocalLLaMA/comments/1keblo7/quick_shoutout_to_qwen330ba3b_as_a_study_tool_for/
| false | false |
self
| 90 | null |
How do you actually run your local LLMs? Trying to map common setups.
| 1 |
[removed]
| 2025-05-04T04:54:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1kebwkf/how_do_you_actually_run_your_local_llms_trying_to/
|
Repulsive_Factor_647
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kebwkf
| false | null |
t3_1kebwkf
|
/r/LocalLLaMA/comments/1kebwkf/how_do_you_actually_run_your_local_llms_trying_to/
| false | false |
self
| 1 | null |
Question about code blocks in Open WebUI
| 1 |
[removed]
| 2025-05-04T06:02:53 |
https://www.reddit.com/r/LocalLLaMA/comments/1kecxcm/question_about_code_blocks_in_open_webui/
|
DrVonSinistro
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kecxcm
| false | null |
t3_1kecxcm
|
/r/LocalLLaMA/comments/1kecxcm/question_about_code_blocks_in_open_webui/
| false | false |
self
| 1 | null |
Vibe-code any AI agent (by Langbase)
| 1 | 2025-05-04T06:12:34 |
https://chai.new
|
AngryBirdenator
|
chai.new
| 1970-01-01T00:00:00 | 0 |
{}
|
1ked2d4
| false | null |
t3_1ked2d4
|
/r/LocalLLaMA/comments/1ked2d4/vibecode_any_ai_agent_by_langbase/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'oenbcRJZQzsWiR13IlNI41FzIt65JIbGOFsrtOpFZfY', 'resolutions': [{'height': 61, 'url': 'https://external-preview.redd.it/XN1EOquhcJqRNzvgrZNynrrtdqf3oLAeq5BqSUZQMVs.jpg?width=108&crop=smart&auto=webp&s=84cb2fad1da70346fce8b10a51cc1610d96a74ec', 'width': 108}, {'height': 123, 'url': 'https://external-preview.redd.it/XN1EOquhcJqRNzvgrZNynrrtdqf3oLAeq5BqSUZQMVs.jpg?width=216&crop=smart&auto=webp&s=ad38811c16b5508f5fb0ca6db18104602e0970ea', 'width': 216}, {'height': 183, 'url': 'https://external-preview.redd.it/XN1EOquhcJqRNzvgrZNynrrtdqf3oLAeq5BqSUZQMVs.jpg?width=320&crop=smart&auto=webp&s=ace3e8aa1d4930a1f7f58f28f19047fbd16792a6', 'width': 320}, {'height': 366, 'url': 'https://external-preview.redd.it/XN1EOquhcJqRNzvgrZNynrrtdqf3oLAeq5BqSUZQMVs.jpg?width=640&crop=smart&auto=webp&s=81eb5b242ab732f0e267679dead5c9f65913d589', 'width': 640}, {'height': 549, 'url': 'https://external-preview.redd.it/XN1EOquhcJqRNzvgrZNynrrtdqf3oLAeq5BqSUZQMVs.jpg?width=960&crop=smart&auto=webp&s=0638831e41c847b9cd4acc28e91e753e467ba02f', 'width': 960}, {'height': 618, 'url': 'https://external-preview.redd.it/XN1EOquhcJqRNzvgrZNynrrtdqf3oLAeq5BqSUZQMVs.jpg?width=1080&crop=smart&auto=webp&s=9c04463260cfdf3987bf55ff26e11974cc34b960', 'width': 1080}], 'source': {'height': 1978, 'url': 'https://external-preview.redd.it/XN1EOquhcJqRNzvgrZNynrrtdqf3oLAeq5BqSUZQMVs.jpg?auto=webp&s=348914e173170ebb39240c32735bb03ad68cc307', 'width': 3456}, 'variants': {}}]}
|
||
How to construct your own evals and learn about evaluations and benchmarking?
| 4 |
Hi!
I'm recruiting for an MLE role for a company which focuses on evals and benchmarking. I suspect that the interviewing process + take-home assessment will focus a lot on these topics (duh), how can I get myself up-to-speed on how to create evals and benchmarks and all that? Sorry for the ambiguous question but any help would be appreciated<3 thank you!!
| 2025-05-04T06:13:57 |
https://www.reddit.com/r/LocalLLaMA/comments/1ked33d/how_to_construct_your_own_evals_and_learn_about/
|
darkGrayAdventurer
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ked33d
| false | null |
t3_1ked33d
|
/r/LocalLLaMA/comments/1ked33d/how_to_construct_your_own_evals_and_learn_about/
| false | false |
self
| 4 | null |
Qwen 3 32b vs QwQ 32b
| 53 |
This is a comparison I barely see and its slightly confusing too as QwQ is kinda a pure reasoning model while Qwen 3 is using reasoning by default but it can be deactivated. In some benchmarks QwQ is even better - so the only advantage of Qwen seems to be that you can use it without reasoning. I assume most benchmarks were done with the default so how good is it without reasoning? Any experience? Other advantages? Or does someone know benchmarks that explicitly test Qwen without reasoning?
| 2025-05-04T06:18:30 |
https://www.reddit.com/r/LocalLLaMA/comments/1ked5iy/qwen_3_32b_vs_qwq_32b/
|
nore_se_kra
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ked5iy
| false | null |
t3_1ked5iy
|
/r/LocalLLaMA/comments/1ked5iy/qwen_3_32b_vs_qwq_32b/
| false | false |
self
| 53 | null |
Ryzen AI Max+ 395 + a gpu?
| 35 |
I see the Ryzen 395 Max+ spec sheet lists 16 PCIe 4.0 lanes. It’s also been use in some desktops. Is there any way to combine a max+ with a cheap 24gb GPU? Like an AMD 7900xtx or a 3090? I feel if you could put shared experts (llama 4) or most frequently used experts (qwen3) on the GPU the 395 max+ would be an absolute beast…
| 2025-05-04T06:31:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1kedbv7/ryzen_ai_max_395_a_gpu/
|
Alarming-Ad8154
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kedbv7
| false | null |
t3_1kedbv7
|
/r/LocalLLaMA/comments/1kedbv7/ryzen_ai_max_395_a_gpu/
| false | false |
self
| 35 | null |
Need a new roof or quick repair? Get free quotes from top-rated local roofers on Thumbtack
| 1 | 2025-05-04T06:38:15 |
https://homesmartprotection.shop/
|
No_Deal_6285
|
homesmartprotection.shop
| 1970-01-01T00:00:00 | 0 |
{}
|
1kedfjl
| false | null |
t3_1kedfjl
|
/r/LocalLLaMA/comments/1kedfjl/need_a_new_roof_or_quick_repair_get_free_quotes/
| false | false |
default
| 1 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.