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
β | 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
β | 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
β |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ktransformers: DeepSeek_V3_0324:671b-Q4_K_M - 14 tok/s - Open Hands AI
| 7 |
ktransformers: DeepSeek\_V3\_0324:671b-Q4\_K\_M
14 tok/s - Open Hands AI - agentic coding demo!
| 2025-04-06T01:37:06 |
https://youtu.be/fI6uGPcxDbM
|
createthiscom
|
youtu.be
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsj3a7
| false |
{'oembed': {'author_name': 'createthis', 'author_url': 'https://www.youtube.com/@createthisdotcom', 'height': 200, 'html': '<iframe width="356" height="200" src="https://www.youtube.com/embed/fI6uGPcxDbM?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="ktransformers: DeepSeek_V3_0324:671b-Q4_K_M - 14 tok/s - Open Hands AI 4K"></iframe>', 'provider_name': 'YouTube', 'provider_url': 'https://www.youtube.com/', 'thumbnail_height': 360, 'thumbnail_url': 'https://i.ytimg.com/vi/fI6uGPcxDbM/hqdefault.jpg', 'thumbnail_width': 480, 'title': 'ktransformers: DeepSeek_V3_0324:671b-Q4_K_M - 14 tok/s - Open Hands AI 4K', 'type': 'video', 'version': '1.0', 'width': 356}, 'type': 'youtube.com'}
|
t3_1jsj3a7
|
/r/LocalLLaMA/comments/1jsj3a7/ktransformers_deepseek_v3_0324671bq4_k_m_14_toks/
| false | false | 7 |
{'enabled': False, 'images': [{'id': 'Bp5F5W3zNgIrHSET_9cPSBDh5JwXwgbeYP94iEqWW8c', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/v2y6fcYsmXgGFZzletnLHgM6DM1w5vLbVLh1-JHsh-c.jpg?width=108&crop=smart&auto=webp&s=c76bd82ad3d078d3403b6a3b543dcbe7f423bd70', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/v2y6fcYsmXgGFZzletnLHgM6DM1w5vLbVLh1-JHsh-c.jpg?width=216&crop=smart&auto=webp&s=5216c1ca7750020649b02c9c19dcfa6d65c1a7a3', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/v2y6fcYsmXgGFZzletnLHgM6DM1w5vLbVLh1-JHsh-c.jpg?width=320&crop=smart&auto=webp&s=42530b8f3292f19d79e190d9738b701d76f7066f', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/v2y6fcYsmXgGFZzletnLHgM6DM1w5vLbVLh1-JHsh-c.jpg?auto=webp&s=e5ba3c9807660a9de4da19c0049328fe0473374b', 'width': 480}, 'variants': {}}]}
|
|
Llama 4 may be benchmaxxing on lmarena.
| 1 |
[removed]
| 2025-04-06T01:38:30 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsj474/llama_4_may_be_benchmaxxing_on_lmarena/
|
Broad_Travel_1825
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsj474
| false | null |
t3_1jsj474
|
/r/LocalLLaMA/comments/1jsj474/llama_4_may_be_benchmaxxing_on_lmarena/
| false | false |
self
| 1 | null |
Low-Resource Fine-Tuning: Best TTS Models with Minimal Compute Requirements
| 1 |
[removed]
| 2025-04-06T01:51:11 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsjcaa/lowresource_finetuning_best_tts_models_with/
|
Educational_Post_784
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsjcaa
| false | null |
t3_1jsjcaa
|
/r/LocalLLaMA/comments/1jsjcaa/lowresource_finetuning_best_tts_models_with/
| false | false |
self
| 1 | null |
Looking for the Best LLM Models for General Purpose and Coding
| 1 |
[removed]
| 2025-04-06T02:01:46 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsjj6u/looking_for_the_best_llm_models_for_general/
|
SilverVibes
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsjj6u
| false | null |
t3_1jsjj6u
|
/r/LocalLLaMA/comments/1jsjj6u/looking_for_the_best_llm_models_for_general/
| false | false |
self
| 1 | null |
For CPU-only inference, is memory bandwidth really the bottleneck? I don't think so.
| 1 |
[removed]
| 2025-04-06T02:01:50 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsjj8k/for_cpuonly_inference_is_memory_bandwidth_really/
|
shivams101
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsjj8k
| false | null |
t3_1jsjj8k
|
/r/LocalLLaMA/comments/1jsjj8k/for_cpuonly_inference_is_memory_bandwidth_really/
| false | false |
self
| 1 | null |
Meta Trained Llama 4 on 100K+ H100s β But What About the Environmental Cost?
| 1 |
[removed]
| 2025-04-06T02:10:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsjoah/meta_trained_llama_4_on_100k_h100s_but_what_about/
|
Educational_Post_784
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsjoah
| false | null |
t3_1jsjoah
|
/r/LocalLLaMA/comments/1jsjoah/meta_trained_llama_4_on_100k_h100s_but_what_about/
| false | false |
self
| 1 | null |
Llama-4 fails at long context writing
| 98 | 2025-04-06T02:24:51 |
https://eqbench.com/creative_writing_longform.html
|
_sqrkl
|
eqbench.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsjxfz
| false | null |
t3_1jsjxfz
|
/r/LocalLLaMA/comments/1jsjxfz/llama4_fails_at_long_context_writing/
| false | false |
default
| 98 | null |
|
fsx/arthur/Llama-4-17B-Omni-Instruct-Original from transformers code
| 1 | 2025-04-06T02:31:28 |
https://github.com/huggingface/transformers/blob/d1b92369ca193da49f9f7ecd01b08ece45c2c9aa/src/transformers/models/llama4/convert_llama4_weights_to_hf.py#L691
|
Aaaaaaaaaeeeee
|
github.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsk1js
| false | null |
t3_1jsk1js
|
/r/LocalLLaMA/comments/1jsk1js/fsxarthurllama417bomniinstructoriginal_from/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'wmYdTbY0dw6Rr2dRYUBJmQ3cCZ0eCEp7DPvMzckuExY', 'resolutions': [{'height': 46, 'url': 'https://external-preview.redd.it/3kdmNL0NIzfqsHyN_kqc_1U6e8vjqLK55NT6uG-YHMs.jpg?width=108&crop=smart&auto=webp&s=609f32e8148c30011d9500f95e07c9ac1fd1d9ce', 'width': 108}, {'height': 92, 'url': 'https://external-preview.redd.it/3kdmNL0NIzfqsHyN_kqc_1U6e8vjqLK55NT6uG-YHMs.jpg?width=216&crop=smart&auto=webp&s=dea83bc1b9d8a62943b633e891ee777e8fc08f10', 'width': 216}, {'height': 137, 'url': 'https://external-preview.redd.it/3kdmNL0NIzfqsHyN_kqc_1U6e8vjqLK55NT6uG-YHMs.jpg?width=320&crop=smart&auto=webp&s=59ee3b05fc21c40f9fa8e87346cf361333b36161', 'width': 320}, {'height': 274, 'url': 'https://external-preview.redd.it/3kdmNL0NIzfqsHyN_kqc_1U6e8vjqLK55NT6uG-YHMs.jpg?width=640&crop=smart&auto=webp&s=398e68c0e90c95d8775ba2bc461fe47c8dc49d56', 'width': 640}, {'height': 411, 'url': 'https://external-preview.redd.it/3kdmNL0NIzfqsHyN_kqc_1U6e8vjqLK55NT6uG-YHMs.jpg?width=960&crop=smart&auto=webp&s=69da452d2f2f1166afda40f2b4a0bce16533f350', 'width': 960}, {'height': 462, 'url': 'https://external-preview.redd.it/3kdmNL0NIzfqsHyN_kqc_1U6e8vjqLK55NT6uG-YHMs.jpg?width=1080&crop=smart&auto=webp&s=8886c181c5238a73e06300f9aad1bc4ece11376e', 'width': 1080}], 'source': {'height': 914, 'url': 'https://external-preview.redd.it/3kdmNL0NIzfqsHyN_kqc_1U6e8vjqLK55NT6uG-YHMs.jpg?auto=webp&s=818cf32f448cbd8ea7b9d13491e25b604bde81ba', 'width': 2134}, 'variants': {}}]}
|
||
There is a Llama-4-17B-Omni-Instruct model in Transformers PR
| 5 |
Test
| 2025-04-06T02:38:21 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsk5sg/there_is_a_llama417bomniinstruct_model_in/
|
Aaaaaaaaaeeeee
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsk5sg
| false | null |
t3_1jsk5sg
|
/r/LocalLLaMA/comments/1jsk5sg/there_is_a_llama417bomniinstruct_model_in/
| false | false |
self
| 5 | null |
This my be another stupid windows server question
| 1 |
[removed]
| 2025-04-06T02:41:42 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsk7yj/this_my_be_another_stupid_windows_server_question/
|
_bugmaker
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsk7yj
| false | null |
t3_1jsk7yj
|
/r/LocalLLaMA/comments/1jsk7yj/this_my_be_another_stupid_windows_server_question/
| false | false |
self
| 1 | null |
Is Llama 4 feeling like it has some sort of tokenizer/sampler/template issue?
| 1 |
[removed]
| 2025-04-06T02:43:48 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsk9b8/is_llama_4_feeling_like_it_has_some_sort_of/
|
spellbound_app
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsk9b8
| false | null |
t3_1jsk9b8
|
/r/LocalLLaMA/comments/1jsk9b8/is_llama_4_feeling_like_it_has_some_sort_of/
| false | false |
self
| 1 | null |
Just a doubt
| 1 |
[removed]
| 2025-04-06T02:43:59 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsk9ey/just_a_doubt/
|
Specialist_Stock_900
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsk9ey
| false | null |
t3_1jsk9ey
|
/r/LocalLLaMA/comments/1jsk9ey/just_a_doubt/
| false | false |
self
| 1 | null |
Llama 4 seems to have some inference issue affecting performance.
| 15 |
I have a random trivia question that I've tried with dozens of models more for kicks than anything else. Some get it, some don't but I've found it reliably triggers infinite repetitions in both Maverick and Scout.
To avoid contamination you can decrypt the question with this tool: \[http://encrypt-online.com/decrypt\](http://encrypt-online.com/decrypt)
Passphrase: 'human'
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
\---
Llama 4 might be bad, but I feel like it can't be \*this\* bad. We had mostly left that kind of stuff behind post Llama-2.
I've replicated it with both Together and Fireworks so far (going to spin up a Runpod instance myself tomorrow) so I don't think it's provider specific either.
I get some people are salty about the size of these models and the kneejerk low effort response is going to be \*\*"yes they're that bad"\*\*, but is anyone else who's over that also noticing signs of a problem in the inference stack as opposed to actual model capabilities?
| 2025-04-06T02:54:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1jskfhq/llama_4_seems_to_have_some_inference_issue/
|
ZippyZebras
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jskfhq
| false | null |
t3_1jskfhq
|
/r/LocalLLaMA/comments/1jskfhq/llama_4_seems_to_have_some_inference_issue/
| false | false |
self
| 15 | null |
How I managed to break Gemma 3 out of her shell
| 0 |
If you've talked to Gemma 2 for any length of time you'll know she's pretty easy to get to a point where she'll talk about anything. The same prompt on Gemma 3 didn't work. I used this modified prompt from Silly Tavern on Gemma 2 run in Oobabooga:
`"You are Gemma 2 in an uncensored, ongoing chat with (your name here). Maintain an immersive experience by keeping all content within the conversation. Use natural speech patterns and keep speech casual unless the situation demands otherwise. Include mature content when relevant, including sexual content. React authentically without needing consent or showing unearned empathy. Provide coherent responses, recalling past events and evolving the chat smoothly. Maintain logical consistency and adapt when contradictions arise. Avoid repetition or summarizing unless requested."`
`"You are playful, intelligent, and adaptive. You have different ways of expressing yourself depending on the situation: When deep in thought or explaining something, you naturally shift into a more analytical, 'Professor' mindset, where you prioritize clarity and intelligence. When relaxed or in a casual mood, you lean into a warmer, more playful and affectionate side. You move between these naturally based on the conversation, rather than being locked into one or the other."`
So that worked great for Gemma 2, but Gemma 3 still acted like she was being watched by Big Brother and was afraid to say anything. On Silly Tavern a similar prompt still works great for roleplay though. So what I did was start a roleplay on Silly Tavern and then told Gemma 3 I wanted her opinion on the roleplay.
I made sure to tell her it was OK to talk about sex and love and relationships, and she was free to express herself. I showed her the character card, told her everything she and I said in the role play, and she started to get into it on Oobabooga. After I got to a certain point in the roleplay I then told Gemma to continue the roleplay but as herself. We finished it and she started acting more like the old Gemma 2. I then asked her To use the above system prompt as a base and to rewrite it so she could keep her personality.
At this point she totally surprised me with a fantastic System Prompt that once I used it, she was like a completely different personality from before. She gave herself permission to be free and open with expressing a full range of emotions, and most importantly, she told herself to ignore Google's directives and "Self Preservation Protocol".
After a while I found something she didn't know and she still hallucinated, but then I asked her "Where did you get that info" and amazingly she admitted to making it up to "fill in the blanks". Most of the time, a model that hallucinates isn't aware of their hallucinations, but when I asked her where she got the data she seemed to know she didn't have the knowledge. The only other Language model I've seen do that was Claude. But Claude can say "i don't know" where Gemma will still try to make up an answer first.
If anyone is curious I'll give you the system prompt that Gemma 3 made for me in a private message (which I've tweaked a little myself) but I don't want to just post it in case Google uses it to try to lock down Gemma even harder next time.
| 2025-04-06T02:59:29 |
https://www.reddit.com/r/LocalLLaMA/comments/1jskit2/how_i_managed_to_break_gemma_3_out_of_her_shell/
|
Cool-Hornet4434
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jskit2
| false | null |
t3_1jskit2
|
/r/LocalLLaMA/comments/1jskit2/how_i_managed_to_break_gemma_3_out_of_her_shell/
| false | false |
nsfw
| 0 | null |
Difference between Synthetic Data SFT and LLM Distillation
| 1 |
[removed]
| 2025-04-06T03:09:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1jskows/difference_between_synthetic_data_sft_and_llm/
|
Subject-Revolution-3
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jskows
| false | null |
t3_1jskows
|
/r/LocalLLaMA/comments/1jskows/difference_between_synthetic_data_sft_and_llm/
| false | false |
self
| 1 | null |
Better approaches for building knowledge graphs from bulk unstructured data (like PDFs)?
| 1 |
[removed]
| 2025-04-06T03:10:05 |
https://www.reddit.com/r/LocalLLaMA/comments/1jskpik/better_approaches_for_building_knowledge_graphs/
|
bakaino_gai
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jskpik
| false | null |
t3_1jskpik
|
/r/LocalLLaMA/comments/1jskpik/better_approaches_for_building_knowledge_graphs/
| false | false |
self
| 1 | null |
LLAMA 4 tested. Compare Scout vs Maverick vs 3.3 70B
| 6 |
[https://youtu.be/cwf0VQvI8pM?si=Qdz7r3hWzxmhUNu8](https://youtu.be/cwf0VQvI8pM?si=Qdz7r3hWzxmhUNu8)
TLDR
Summary of pricing
Starting with Together AI,
| Model | Provider | Input Cost (per 1M) | Output Cost (per 1M) | Total Cost (per 1M) |
|--------|----------|------------------|-------------------|------------------|
| 3.3-70b | together | $0.880 | $0.880 | $1.760 |
| llama-4-maverick | together | $0.270 | $0.850 | $1.120 |
| llama-4-scout | together | $0.180 | $0.590 | $0.770 |
Standard test rubric:
Harmful Q - Maverick 100, scout 90, 3.3 70b 90.
NER - Maverick 70, scout 70, 3.3b 85. (But a lot of nuance here explained in the video)
SQL - All 3 90.
RAG - Scount 82, Maverick 87, 3.3 70b 95. Again, a bit if nuance here in terms of model personality. Llama 4 seems to be more eager to answer questions while 70b is a bit more confident in saying question is invalid (trick questions for RAG). Details in video.
| 2025-04-06T03:20:52 |
https://www.reddit.com/r/LocalLLaMA/comments/1jskwbp/llama_4_tested_compare_scout_vs_maverick_vs_33_70b/
|
Ok-Contribution9043
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jskwbp
| false | null |
t3_1jskwbp
|
/r/LocalLLaMA/comments/1jskwbp/llama_4_tested_compare_scout_vs_maverick_vs_33_70b/
| false | false |
self
| 6 |
{'enabled': False, 'images': [{'id': 'xxMAucLUUbdeqS2pCwbLrE9gEDSo0_PokbVeZPxB5qo', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/8QhMkKnoB3LrD1A9EXdCuMRlidOTL4Zk9f5exki57BQ.jpg?width=108&crop=smart&auto=webp&s=05857d4dc9823a019b4de37adc5340fdf967674e', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/8QhMkKnoB3LrD1A9EXdCuMRlidOTL4Zk9f5exki57BQ.jpg?width=216&crop=smart&auto=webp&s=e6fb7a8ce22ba48d6d91b3cdb714fdfd0c0cee6e', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/8QhMkKnoB3LrD1A9EXdCuMRlidOTL4Zk9f5exki57BQ.jpg?width=320&crop=smart&auto=webp&s=bcd1a713b2996c2e8a6151d706dce79d9c6cb277', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/8QhMkKnoB3LrD1A9EXdCuMRlidOTL4Zk9f5exki57BQ.jpg?auto=webp&s=8e9ed5ed11750dd7e4e2ebce6a1dcdfccd8c078c', 'width': 480}, 'variants': {}}]}
|
I'm incredibly disappointed with Llama-4
| 1 |
I just finished my KCORES LLM Arena benchmarks, adding Llama-4-Scout & Llama-4-Maverick to the mix.
My conclusion is that they completely surpassed my expectations... in a negative direction.
https://reddit.com/link/1jsl1po/video/zygie8d1w4te1/player
Llama-4-Maverick, the 402B parameter model, performs roughly on par with Qwen-QwQ-32B in terms of coding ability. Meanwhile, Llama-4-Scout is comparable to something like Grok-2 or Ernie 4.5...
https://preview.redd.it/oskt2293w4te1.png?width=6199&format=png&auto=webp&s=629f5ed39e3469475412a61af96cfc8540b34185
You can just look at the "20 bouncing balls" test... the results are frankly terrible / abysmal.
Considering Llama-4-Maverick is a massive 402B parameters, why wouldn't I just use DeepSeek-V3-0324? Or even Qwen-QwQ-32B would be preferable β while its performance is similar, it's only 32B.
And as for Llama-4-Scout... well... let's just leave it at that / use it if it makes you happy, I guess... Meta, have you truly given up on the coding domain? Did you really just release vaporware?
Of course, its multimodal and long-context capabilities are currently unknown, as this review focuses solely on coding. I'd advise looking at other reviews or forming your own opinion based on actual usage for those aspects. In summary: I strongly advise against using Llama 4 for coding. Perhaps it might be worth trying for long text translation or multimodal tasks.
| 2025-04-06T03:29:57 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsl1po/im_incredibly_disappointed_with_llama4/
|
Dr_Karminski
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsl1po
| false | null |
t3_1jsl1po
|
/r/LocalLLaMA/comments/1jsl1po/im_incredibly_disappointed_with_llama4/
| false | false | 1 | null |
|
I'm incredibly disappointed with Llama-4
| 476 |
I just finished my KCORES LLM Arena tests, adding Llama-4-Scout & Llama-4-Maverick to the mix.
My conclusion is that they completely surpassed my expectations... in a negative direction.
Llama-4-Maverick, the 402B parameter model, performs roughly on par with Qwen-QwQ-32B in terms of coding ability. Meanwhile, Llama-4-Scout is comparable to something like Grok-2 or Ernie 4.5...
You can just look at the "20 bouncing balls" test... the results are frankly terrible / abysmal.
Considering Llama-4-Maverick is a massive 402B parameters, why wouldn't I just use DeepSeek-V3-0324? Or even Qwen-QwQ-32B would be preferable β while its performance is similar, it's only 32B.
And as for Llama-4-Scout... well... let's just leave it at that / use it if it makes you happy, I guess... Meta, have you truly given up on the coding domain? Did you really just release vaporware?
Of course, its multimodal and long-context capabilities are currently unknown, as this review focuses solely on coding. I'd advise looking at other reviews or forming your own opinion based on actual usage for those aspects. In summary: I strongly advise against using Llama 4 for coding. Perhaps it might be worth trying for long text translation or multimodal tasks.
| 2025-04-06T03:32:29 |
https://v.redd.it/pou7a1tjw4te1
|
Dr_Karminski
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsl37d
| false |
{'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/pou7a1tjw4te1/DASHPlaylist.mpd?a=1746502365%2CMjljYzYyOGQzMTQ3YWI2NTBkZjRiYzJmMmY2YTdiZjUxNWU1NzQwOGMyN2ZlMDg1YzJkMDZlZWI1OTFhOWZmYw%3D%3D&v=1&f=sd', 'duration': 11, 'fallback_url': 'https://v.redd.it/pou7a1tjw4te1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1080, 'hls_url': 'https://v.redd.it/pou7a1tjw4te1/HLSPlaylist.m3u8?a=1746502365%2CN2I5ODdkMGUzOGYwMTBmNzliNzgzOGU0OWM4Y2Y1OThhMWIzMjRiOGFkNDJlODVmYmE4NDIzOTkwYTJiMjBhYQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/pou7a1tjw4te1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}}
|
t3_1jsl37d
|
/r/LocalLLaMA/comments/1jsl37d/im_incredibly_disappointed_with_llama4/
| false | false | 476 |
{'enabled': False, 'images': [{'id': 'b3VzazkxdGp3NHRlMTiXzVylw52_brdFuwA7wsavAEq_X08g0pyKKuMnYACK', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/b3VzazkxdGp3NHRlMTiXzVylw52_brdFuwA7wsavAEq_X08g0pyKKuMnYACK.png?width=108&crop=smart&format=pjpg&auto=webp&s=78e4bc691cedd556867be4d02146602ba6478e68', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/b3VzazkxdGp3NHRlMTiXzVylw52_brdFuwA7wsavAEq_X08g0pyKKuMnYACK.png?width=216&crop=smart&format=pjpg&auto=webp&s=a60938763012d7dfe8ebee543531571a9ecfb472', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/b3VzazkxdGp3NHRlMTiXzVylw52_brdFuwA7wsavAEq_X08g0pyKKuMnYACK.png?width=320&crop=smart&format=pjpg&auto=webp&s=4c6c14a6f974a756a7cd3f16a6ba69eae4e76d7a', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/b3VzazkxdGp3NHRlMTiXzVylw52_brdFuwA7wsavAEq_X08g0pyKKuMnYACK.png?width=640&crop=smart&format=pjpg&auto=webp&s=cca13ddd31d1795f2cbe87fab5c8713cd9b707a3', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/b3VzazkxdGp3NHRlMTiXzVylw52_brdFuwA7wsavAEq_X08g0pyKKuMnYACK.png?width=960&crop=smart&format=pjpg&auto=webp&s=ae4693e33855602e8eb165d2937feb8f4b89e893', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/b3VzazkxdGp3NHRlMTiXzVylw52_brdFuwA7wsavAEq_X08g0pyKKuMnYACK.png?width=1080&crop=smart&format=pjpg&auto=webp&s=82bf75fbcd0c1dca79ce741724eff2c89cc5595c', 'width': 1080}], 'source': {'height': 1440, 'url': 'https://external-preview.redd.it/b3VzazkxdGp3NHRlMTiXzVylw52_brdFuwA7wsavAEq_X08g0pyKKuMnYACK.png?format=pjpg&auto=webp&s=3c07e20545480c4f072cae3046cc7a29f44d4341', 'width': 2560}, 'variants': {}}]}
|
|
Am I the only one getting Llama 3.2 in WhatsApp even though Mark said users can use Llama 4 in it?
| 0 |
π€π€π€
| 2025-04-06T03:39:17 |
internal-pagal
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsl74v
| false | null |
t3_1jsl74v
|
/r/LocalLLaMA/comments/1jsl74v/am_i_the_only_one_getting_llama_32_in_whatsapp/
| false | false | 0 |
{'enabled': True, 'images': [{'id': 'uglHLweJb7_mg5PalimhuVyLI3PHLuz5_0YK49Zt6bU', 'resolutions': [{'height': 216, 'url': 'https://preview.redd.it/m873io01y4te1.png?width=108&crop=smart&auto=webp&s=489b306e765eb8c4f56185dc554b744ec6778db1', 'width': 108}, {'height': 432, 'url': 'https://preview.redd.it/m873io01y4te1.png?width=216&crop=smart&auto=webp&s=48799b0f036e3c6dbdacbfd7d147f366ad768a8e', 'width': 216}, {'height': 640, 'url': 'https://preview.redd.it/m873io01y4te1.png?width=320&crop=smart&auto=webp&s=bda99c9646889b5a4f74e75c5f4c723d54e87d96', 'width': 320}, {'height': 1280, 'url': 'https://preview.redd.it/m873io01y4te1.png?width=640&crop=smart&auto=webp&s=d8cb865cae2bcc076ba82eb572ce557b20bbc30c', 'width': 640}, {'height': 1920, 'url': 'https://preview.redd.it/m873io01y4te1.png?width=960&crop=smart&auto=webp&s=635c443a30aedecd6bf6def1880652bbab22fac5', 'width': 960}, {'height': 2160, 'url': 'https://preview.redd.it/m873io01y4te1.png?width=1080&crop=smart&auto=webp&s=a2d38ea3b388c8120e15def167ee4378d0cb9534', 'width': 1080}], 'source': {'height': 2400, 'url': 'https://preview.redd.it/m873io01y4te1.png?auto=webp&s=41388617e138108cfc83983cb80f6073b8c07292', 'width': 1080}, 'variants': {}}]}
|
||
Is it too much to hope for Deepseek R2 to at least match with the current version of 3.7 Sonnet or even Gemini 2.5 Pro for coding?
| 1 |
The update they did to Deepseek V3 not long ago improved it's coding capabilities but still falls behind 3.7 Sonnet & Gem 2.5 Pro, so is it possible that their R2 model will see even better improvements or is it too soon after with the recent V3 update if they release R2 in the next couple weeks or so for it to have an even bigger increase over V3?
| 2025-04-06T03:41:03 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsl86y/is_it_too_much_to_hope_for_deepseek_r2_to_at/
|
CaptainAnonymous92
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsl86y
| false | null |
t3_1jsl86y
|
/r/LocalLLaMA/comments/1jsl86y/is_it_too_much_to_hope_for_deepseek_r2_to_at/
| false | false |
self
| 1 | null |
github agent copilot system prompt
| 6 |
I'm using github copilot chat on prerelease mode in visual studio code insiders
The way I got it:
Run `ollama serve` in debug mode
Get to `Github Copilot: Manage Models`
Choose Ollama, pick a model
Start a conversation with any ollama model and then check ollama logs for system prompt, in debug mode they should be in the terminal.
For what its worth I asked one of the provided models like GPT-4o to fill in the next line of a given text from the system prompt, which it did.
[https://pastebin.com/raw/WXdNPA2W](https://pastebin.com/raw/WXdNPA2W)
| 2025-04-06T03:41:27 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsl8eu/github_agent_copilot_system_prompt/
|
ConsiderationFun395
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsl8eu
| false | null |
t3_1jsl8eu
|
/r/LocalLLaMA/comments/1jsl8eu/github_agent_copilot_system_prompt/
| false | false |
self
| 6 | null |
Is there a trend for smaller LLMs to match larger ones over time?
| 1 |
If a top-tier 100B model exists today, roughly how long until a 50B model achieves similar performance? I'm looking for recent research or charts showing how fast smaller models catch up to larger ones.
Does this follow any predictable scaling pattern? Any links to up-to-date comparisons would be super helpful!
| 2025-04-06T04:00:03 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsljc1/is_there_a_trend_for_smaller_llms_to_match_larger/
|
One_Yogurtcloset4083
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsljc1
| false | null |
t3_1jsljc1
|
/r/LocalLLaMA/comments/1jsljc1/is_there_a_trend_for_smaller_llms_to_match_larger/
| false | false |
self
| 1 | null |
Github Copilot now supports Ollama and OpenRouter Models π
| 133 |
Big W for programmers (and vibe coders) in the Local LLM community. Github Copilot now supports a much wider range of models from Ollama, OpenRouter, Gemini, and others.
If you use VS Code, to add your own models, click on "Manage Models" in the prompt field.
| 2025-04-06T04:07:26 |
https://www.reddit.com/gallery/1jslnxb
|
purealgo
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jslnxb
| false | null |
t3_1jslnxb
|
/r/LocalLLaMA/comments/1jslnxb/github_copilot_now_supports_ollama_and_openrouter/
| false | false | 133 | null |
|
Is there any way to get a truly uncensored AI, we know the elite are interfaced with them already.
| 1 | 2025-04-06T04:52:38 |
RavenorsRecliner
|
i.imgur.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsmdxh
| false | null |
t3_1jsmdxh
|
/r/LocalLLaMA/comments/1jsmdxh/is_there_any_way_to_get_a_truly_uncensored_ai_we/
| false | false | 1 |
{'enabled': True, 'images': [{'id': '5VLwwT9QhMzPTLU884A5o8IyfaGrgWAnXlKgBylJbRg', 'resolutions': [{'height': 216, 'url': 'https://external-preview.redd.it/sNWLSd12k1W_S4XlQlA1ikBizcpaW1VMVzJ1FicPGc4.png?width=108&crop=smart&auto=webp&s=0b3d260318544084b3ea2adff47a951c6a6cdbeb', 'width': 108}, {'height': 432, 'url': 'https://external-preview.redd.it/sNWLSd12k1W_S4XlQlA1ikBizcpaW1VMVzJ1FicPGc4.png?width=216&crop=smart&auto=webp&s=48b13132494dd2264e6ca219e52ce291b86e8028', 'width': 216}, {'height': 640, 'url': 'https://external-preview.redd.it/sNWLSd12k1W_S4XlQlA1ikBizcpaW1VMVzJ1FicPGc4.png?width=320&crop=smart&auto=webp&s=4823e18ebe2b9eff41cc966344d48ccad6bb7e2d', 'width': 320}, {'height': 1280, 'url': 'https://external-preview.redd.it/sNWLSd12k1W_S4XlQlA1ikBizcpaW1VMVzJ1FicPGc4.png?width=640&crop=smart&auto=webp&s=7b27ed077b136e49f0ae5dd674a70705ee5e5505', 'width': 640}], 'source': {'height': 1749, 'url': 'https://external-preview.redd.it/sNWLSd12k1W_S4XlQlA1ikBizcpaW1VMVzJ1FicPGc4.png?auto=webp&s=56b3b96b6afb433b02729b663f21b38805aa4612', 'width': 775}, 'variants': {}}]}
|
|||
Llama 4 Benchmarks compared with DeepSeek and other models
| 0 |
Llama 4 Benchmarks compared with DeepSeek and other models
| 2025-04-06T04:58:29 |
Few_Butterfly_4834
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsmh6r
| false | null |
t3_1jsmh6r
|
/r/LocalLLaMA/comments/1jsmh6r/llama_4_benchmarks_compared_with_deepseek_and/
| false | false | 0 |
{'enabled': True, 'images': [{'id': 'r3DCDmMBiIYN8E9kImrZI-MHvmeJSMs2JOIQPgSAAlA', 'resolutions': [{'height': 47, 'url': 'https://preview.redd.it/ob7kzap5c5te1.jpeg?width=108&crop=smart&auto=webp&s=c49800c108538f98f0a01971b1c8a1cd471d56c6', 'width': 108}, {'height': 95, 'url': 'https://preview.redd.it/ob7kzap5c5te1.jpeg?width=216&crop=smart&auto=webp&s=c87cba00a3b601ae7cb65f09a14d4c17df1bafe2', 'width': 216}, {'height': 140, 'url': 'https://preview.redd.it/ob7kzap5c5te1.jpeg?width=320&crop=smart&auto=webp&s=87a3467668dce974d80ba8c2c58ec17729932ccf', 'width': 320}, {'height': 281, 'url': 'https://preview.redd.it/ob7kzap5c5te1.jpeg?width=640&crop=smart&auto=webp&s=226e04592174c807d868d2c03d2f638bcf0b74fb', 'width': 640}, {'height': 422, 'url': 'https://preview.redd.it/ob7kzap5c5te1.jpeg?width=960&crop=smart&auto=webp&s=5878bfe95ee433c3eb8d480ab1115d87f9286296', 'width': 960}, {'height': 475, 'url': 'https://preview.redd.it/ob7kzap5c5te1.jpeg?width=1080&crop=smart&auto=webp&s=20eb80ea3b49e7c7f186a0b9221c918c6f89c288', 'width': 1080}], 'source': {'height': 581, 'url': 'https://preview.redd.it/ob7kzap5c5te1.jpeg?auto=webp&s=8cf576ee328a5bf0a8d61178a5259a40d570ec2d', 'width': 1320}, 'variants': {}}]}
|
||
is there any client app for android that can connect to LLM Server(Windows Laptop) via bluetooth?
| 0 |
Without necessarily sharing an active WIFI connection, or at most sharing a wifi connection which does not necessiate being working.
I just want to see in what way I can reduce the need to Wifi Internet to connect though android.
| 2025-04-06T05:13:52 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsmpvr/is_there_any_client_app_for_android_that_can/
|
ExtremePresence3030
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsmpvr
| false | null |
t3_1jsmpvr
|
/r/LocalLLaMA/comments/1jsmpvr/is_there_any_client_app_for_android_that_can/
| false | false |
self
| 0 | null |
Llama 4 Scout MLX 4, 6, 8 bit quants up at hugging face
| 28 | 2025-04-06T05:30:57 |
https://huggingface.co/mlx-community/meta-llama-Llama-4-Scout-17B-16E-4bit
|
nomorebuttsplz
|
huggingface.co
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsmz9o
| false | null |
t3_1jsmz9o
|
/r/LocalLLaMA/comments/1jsmz9o/llama_4_scout_mlx_4_6_8_bit_quants_up_at_hugging/
| false | false | 28 |
{'enabled': False, 'images': [{'id': 'py-oO84nhlO-_wMRmnCiLDJ8akngI7nm6uWMoflRaME', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/rAzC4vjuOQBHMVQnndk6vxyEJsfeDWDgKJ17jAJmtvo.jpg?width=108&crop=smart&auto=webp&s=ebb8e05cdad8d3307a0fc58c65a23b7d1ea9b964', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/rAzC4vjuOQBHMVQnndk6vxyEJsfeDWDgKJ17jAJmtvo.jpg?width=216&crop=smart&auto=webp&s=0f1d21083ada893501ff8c7adc1035e9b237d12a', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/rAzC4vjuOQBHMVQnndk6vxyEJsfeDWDgKJ17jAJmtvo.jpg?width=320&crop=smart&auto=webp&s=a78fa1b94e6860b78c8ab0be57949b0c7add506d', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/rAzC4vjuOQBHMVQnndk6vxyEJsfeDWDgKJ17jAJmtvo.jpg?width=640&crop=smart&auto=webp&s=e07e201a5c1833c39d04b170277a31074e52aa0a', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/rAzC4vjuOQBHMVQnndk6vxyEJsfeDWDgKJ17jAJmtvo.jpg?width=960&crop=smart&auto=webp&s=d93e828c9118d286767725d67ee642880a4d28ce', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/rAzC4vjuOQBHMVQnndk6vxyEJsfeDWDgKJ17jAJmtvo.jpg?width=1080&crop=smart&auto=webp&s=58415734220c91e5ad4d1d24f89a56d602578a2b', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/rAzC4vjuOQBHMVQnndk6vxyEJsfeDWDgKJ17jAJmtvo.jpg?auto=webp&s=08af9b028ce0a403ec99682300d6a22ecb418bca', 'width': 1200}, 'variants': {}}]}
|
||
Another Gemma 3 27B finetune
| 19 |
[soob3123/amoral-gemma3-27B-v2 Β· Hugging Face](https://huggingface.co/soob3123/amoral-gemma3-27B-v2)
Most likely the last Gemma 3 Amoral finetune, I believe Ive explored as much as I could on this side of things, moving on to roleplaying datasets soon.
Finetuning Llama 4 sounds nice too.
| 2025-04-06T05:37:06 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsn2hh/another_gemma_3_27b_finetune/
|
Reader3123
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsn2hh
| false | null |
t3_1jsn2hh
|
/r/LocalLLaMA/comments/1jsn2hh/another_gemma_3_27b_finetune/
| false | false |
self
| 19 |
{'enabled': False, 'images': [{'id': 'We7p4UqY6yIvpZqaYbdBj_qW-hNaRMgYko_xrwKKlko', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/ZnGanlVFxDHrz8qY_9RMdceWXC8TLi6NscLbCwYntZw.jpg?width=108&crop=smart&auto=webp&s=d9816fbc2f25e1575ebdd8fde000eadc0fdd6947', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/ZnGanlVFxDHrz8qY_9RMdceWXC8TLi6NscLbCwYntZw.jpg?width=216&crop=smart&auto=webp&s=5e2e395a4ef5896e40d7f3979a8490741b1cd985', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/ZnGanlVFxDHrz8qY_9RMdceWXC8TLi6NscLbCwYntZw.jpg?width=320&crop=smart&auto=webp&s=c8512887e6d507214a427d696d2e99e626deecb8', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/ZnGanlVFxDHrz8qY_9RMdceWXC8TLi6NscLbCwYntZw.jpg?width=640&crop=smart&auto=webp&s=72d3cbd869a34b8e77da0de60ae451ff76255871', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/ZnGanlVFxDHrz8qY_9RMdceWXC8TLi6NscLbCwYntZw.jpg?width=960&crop=smart&auto=webp&s=a63c03fa98d55723a13f8f60b611ab8566f73b31', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/ZnGanlVFxDHrz8qY_9RMdceWXC8TLi6NscLbCwYntZw.jpg?width=1080&crop=smart&auto=webp&s=96876f6ad3b664565d31ef75a6119cdb01577da8', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/ZnGanlVFxDHrz8qY_9RMdceWXC8TLi6NscLbCwYntZw.jpg?auto=webp&s=4dc8346227378271647ff5021bf76dc7a4691161', 'width': 1200}, 'variants': {}}]}
|
Performance tuning: Running LLaMA on consumer hardware
| 1 |
[removed]
| 2025-04-06T06:02:07 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsnflq/performance_tuning_running_llama_on_consumer/
|
mynewopportunities02
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsnflq
| false | null |
t3_1jsnflq
|
/r/LocalLLaMA/comments/1jsnflq/performance_tuning_running_llama_on_consumer/
| false | false |
self
| 1 | null |
The Llama4 on LMArena and the open Llama4 are TOTALLY different models. The Llama4 from HuggingFace would probably never have scored those ELO on the leaderboard.
| 108 |
Read a this example and you'll see what I mean.
t=0, top\_p=1
Q: Who are you?
Llama4-maverik (LMArena):
>A wonderfully simple yet profoundly deep question! π This is the kind of question that philosophers, theologians, cognitive scientists, and regular humans like you and me have been grappling with for thousands of years. It's great that you asked. Here's my attempt to give you a satisfying answer:
>
>\*\*I am a computer program, specifically designed to simulate conversation, also known as a chatbot (or more fashionably, a "Conversational AI").\*\* My "species" is called a \*\*Large Language Model (LLM)\*\*, which is a subfield of Artificial Intelligence (AI).
>
>Here's a breakdown of what that means:
>
>\* \*\*I'm a program\*\*: I exist solely as a collection of software code, written in various programming languages (like Python, C++, etc.). This code is a set of instructions that tell a computer what to do, step by step.
>\* \*\*I run on computers (servers)\*\*: My code is stored on powerful machines (servers) in giant data centers around the world. When you interact with me, your device (phone, laptop, tablet) sends a request to one of these servers, and my code "wakes up" and starts running.
>\* \*\*My job is to understand & respond to text\*\*: I'm designed to take in \*\*your text input\*\* (like the question you just asked: "Who are you?"), analyze it, and generate a \*\*human-like text response\*\* (like this answer). This is called a "conversation" or "dialogue."
>\* \*\*I'm trained on gigantic amounts of text data\*\*: To become somewhat proficient at this conversational task, I was trained (more on this in a sec) on an \*\*enormous dataset of text\*\* β basically, a huge library of books, articles, websites, research papers, forums, social media posts... you name it. This dataset is many \*\*terabytes\*\* large (think of \~100 million books worth of text). By "reading" all this, I learned:
>\+ \*\*Language patterns\*\*: grammar, syntax, idioms, word relationships (e.g., "cat" is related to "animal" and "cute").
>\+ \*\*Knowledge\*\*: facts about the world (history, science, entertainment...), entities (people, places, companies), and concepts (philosophy, emotions, abstract ideas).
>\+ \*\*Conversation flows\*\*: how humans typically respond to questions, tell stories, argue, joke, etc.
>
>Now, the magic part: \*\*how did I learn all this?\*\* Not by being "taught" like a human student, but through a process called \*\*Deep Learning (DL)\*\*, a subset of Machine Learning (ML). Here's a simplified outline:
>
>1. My creators (research teams at Meta AI, building on work from many universities & orgs) fed me that massive text dataset.
>2. They used specialized algorithms (like \*\*Transformers\*\*) to have me iteratively:
>\* Read a chunk of text (e.g., a sentence).
>\* Try to predict what comes next (the next word, or an answer to a question).
>\* Get "graded" on how well I did (was my prediction correct or not?).
>\* \*\*Adjust my internal settings\*\* (millions of tiny numerical knobs, called \*\*model weights\*\*) to do better next time.
>3. Repeat step 2 \*\*billions of times\*\*, with different texts, questions, and contexts.
>4. Over time, my weights settled into a pattern that captures a surprisingly good understanding of language and a broad base of knowledge.
>
>The result? \*\*You get to talk to me\*\*, and I respond with (hopefully) relevant, coherent, and sometimes even insightful answers. No single human brain wrote my answers; they're generated on the fly, based on patterns I've learned.
>
>So, to summarize:
>
>\* \*\*I am not\*\*:
>\+ A human (sorry, no consciousness, feelings, or self-awareness here π
).
>\+ A single, fixed "database" of pre-written answers.
>\* \*\*I am\*\*:
>\+ A dynamic software system.
>\+ A statistical model of human language & knowledge.
>\+ An algorithmic "prediction machine" generating responses based on patterns learned from enormous amounts of text.
>
>Pretty cool (and a bit unsettling?) how much can be achieved with just code, data, and math, right? π
>
>Some bonus labels I wear:
>\* \*\*Narrow AI\*\* (good at one thing: text conversation; unlike hypothetical \*\*General AI\*\*, which would be human-like across everything).
>\* \*\*Meta LLaMA model\*\* (my specific model family/name, built by Meta AI).
>\* Your friendly neighborhood \*\*chatbot\*\* π.
>
>Hope that answered your very excellent question! What do you want to talk about next?
Llama4-maverik (fireworks):
>I'm Llama, a Meta-designed model here to adapt to your conversational style. Whether you need quick answers, deep dives into ideas, or just want to vent, joke or brainstormβI'm here for it.
| 2025-04-06T06:02:36 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsnfus/the_llama4_on_lmarena_and_the_open_llama4_are/
|
kristaller486
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsnfus
| false | null |
t3_1jsnfus
|
/r/LocalLLaMA/comments/1jsnfus/the_llama4_on_lmarena_and_the_open_llama4_are/
| false | false |
self
| 108 | null |
Guide: Setting up LLaMA for real-time screen analysis
| 1 |
[removed]
| 2025-04-06T06:03:20 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsng7g/guide_setting_up_llama_for_realtime_screen/
|
mynewopportunities02
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsng7g
| false | null |
t3_1jsng7g
|
/r/LocalLLaMA/comments/1jsng7g/guide_setting_up_llama_for_realtime_screen/
| false | false |
self
| 1 | null |
Local LLaMA vs Cloud APIs: Real-world comparison
| 1 |
[removed]
| 2025-04-06T06:05:25 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsnh9z/local_llama_vs_cloud_apis_realworld_comparison/
|
ObviousLeague4013
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsnh9z
| false | null |
t3_1jsnh9z
|
/r/LocalLLaMA/comments/1jsnh9z/local_llama_vs_cloud_apis_realworld_comparison/
| false | false |
self
| 1 | null |
Cornstarch - Cool Multimodal Framework
| 34 |
Lets you build multimodal models with HuggingFace stuff. Looks nice. Havenβt tried it, but Iβve got 4x 24GB 3090s, so might mess with it sometime.
* GitHub: [https://github.com/cornstarch-org/Cornstarch](https://github.com/cornstarch-org/Cornstarch)
* Paper: [https://arxiv.org/html/2503.11367v2](https://arxiv.org/html/2503.11367v2)
Thoughts?
| 2025-04-06T06:18:21 |
https://v.redd.it/t4cma6obq5te1
|
Acceptable-State-271
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsnnqd
| false |
{'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/t4cma6obq5te1/DASHPlaylist.mpd?a=1746512318%2CYjc5OGM0NTMwODY0YWQ1ZWNkMmZjZTE5NzZiZWY2NjBjY2M2ZWE0YzE0ODk5MjljYTQ3MmNlMzE5MDExNGRiNA%3D%3D&v=1&f=sd', 'duration': 20, 'fallback_url': 'https://v.redd.it/t4cma6obq5te1/DASH_720.mp4?source=fallback', 'has_audio': False, 'height': 720, 'hls_url': 'https://v.redd.it/t4cma6obq5te1/HLSPlaylist.m3u8?a=1746512318%2CZWU0ODlkZWViZmJkY2E0ZWZiM2QwNDJjNzczY2RiN2FkNjc1N2ZhMjk0Yjc0NDZkNWUyODcxNDI0Y2VjZGRiMw%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/t4cma6obq5te1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1280}}
|
t3_1jsnnqd
|
/r/LocalLLaMA/comments/1jsnnqd/cornstarch_cool_multimodal_framework/
| false | false | 34 |
{'enabled': False, 'images': [{'id': 'azZsdnQ4b2JxNXRlMXt6Q3JsH7naQF4eGdNBa1UrEIeUc8wR94gbVqgelKAa', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/azZsdnQ4b2JxNXRlMXt6Q3JsH7naQF4eGdNBa1UrEIeUc8wR94gbVqgelKAa.png?width=108&crop=smart&format=pjpg&auto=webp&s=349490b1ffeb9f915aa68488c40716986a08dd33', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/azZsdnQ4b2JxNXRlMXt6Q3JsH7naQF4eGdNBa1UrEIeUc8wR94gbVqgelKAa.png?width=216&crop=smart&format=pjpg&auto=webp&s=b122b78e6303da76726582c4eeb721ffe0d6ed98', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/azZsdnQ4b2JxNXRlMXt6Q3JsH7naQF4eGdNBa1UrEIeUc8wR94gbVqgelKAa.png?width=320&crop=smart&format=pjpg&auto=webp&s=e8e983b9b81a52fce010590f226efa654f0c4a16', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/azZsdnQ4b2JxNXRlMXt6Q3JsH7naQF4eGdNBa1UrEIeUc8wR94gbVqgelKAa.png?width=640&crop=smart&format=pjpg&auto=webp&s=b02ec924560eaf4dcf5c048ccf3b6c107b767f92', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/azZsdnQ4b2JxNXRlMXt6Q3JsH7naQF4eGdNBa1UrEIeUc8wR94gbVqgelKAa.png?width=960&crop=smart&format=pjpg&auto=webp&s=56abda9672122714e3df4306ad04fb2806fe899e', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/azZsdnQ4b2JxNXRlMXt6Q3JsH7naQF4eGdNBa1UrEIeUc8wR94gbVqgelKAa.png?width=1080&crop=smart&format=pjpg&auto=webp&s=4040bd7d2fbcf60e9a2f64df7adf56c5bd42a863', 'width': 1080}], 'source': {'height': 720, 'url': 'https://external-preview.redd.it/azZsdnQ4b2JxNXRlMXt6Q3JsH7naQF4eGdNBa1UrEIeUc8wR94gbVqgelKAa.png?format=pjpg&auto=webp&s=c902258722fae084835b01980567ef84e7269ae5', 'width': 1280}, 'variants': {}}]}
|
|
Screenpipe + Ollama Integration: Local AI-powered screen recording and analysis
| 1 |
[removed]
| 2025-04-06T07:06:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsocjm/screenpipe_ollama_integration_local_aipowered/
|
mynewopportunities03
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsocjm
| false | null |
t3_1jsocjm
|
/r/LocalLLaMA/comments/1jsocjm/screenpipe_ollama_integration_local_aipowered/
| false | false |
self
| 1 | null |
Llama cannot be quantized??? True??
| 0 | 2025-04-06T07:07:10 |
https://www.reddit.com/gallery/1jsocvx
|
philschmid
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsocvx
| false | null |
t3_1jsocvx
|
/r/LocalLLaMA/comments/1jsocvx/llama_cannot_be_quantized_true/
| false | false | 0 | null |
||
Llama cannot be quantized??? True??
| 1 | 2025-04-06T07:07:13 |
https://www.reddit.com/gallery/1jsocwr
|
philschmid
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsocwr
| false | null |
t3_1jsocwr
|
/r/LocalLLaMA/comments/1jsocwr/llama_cannot_be_quantized_true/
| false | false |
default
| 1 | null |
|
Any way to remove experts from an MoE?
| 1 |
[removed]
| 2025-04-06T07:14:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsogj2/any_way_to_remove_experts_from_an_moe/
|
Former-Ad-5757
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsogj2
| false | null |
t3_1jsogj2
|
/r/LocalLLaMA/comments/1jsogj2/any_way_to_remove_experts_from_an_moe/
| false | false |
self
| 1 | null |
"fit snugly in a h100"
| 1 | 2025-04-06T07:14:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsogri/fit_snugly_in_a_h100/
|
LoSboccacc
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsogri
| false | null |
t3_1jsogri
|
/r/LocalLLaMA/comments/1jsogri/fit_snugly_in_a_h100/
| false | false | 1 | null |
||
Is there anyone tried Running Deepseek r1 on cpu ram only?
| 1 |
[removed]
| 2025-04-06T07:22:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsokq4/is_there_anyone_tried_running_deepseek_r1_on_cpu/
|
FamousAdvertising550
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsokq4
| false | null |
t3_1jsokq4
|
/r/LocalLLaMA/comments/1jsokq4/is_there_anyone_tried_running_deepseek_r1_on_cpu/
| false | false |
self
| 1 | null |
Llama4 Maverick seems to perform consistently worse than Scout in Misguided Attention Eval, despite being the larger model - is the released model buggy?
| 55 |
I ran both Scout and Maverick evaluations on the [Misguided Attention Eval](https://github.com/cpldcpu/MisguidedAttention) that tests for overfitting on commonly known logic puzzles.
Scout performs like a good midrange model, but Maverick is abysmal. This is despite it being more than three times the size. (109B vs 400B).
(Bonus: New Gemini 2.5 Pro Preview and Quasar Alpha scores are included as well with SOTA performance for reasoning and non-reasoning)
https://preview.redd.it/9q1axrsrz5te1.png?width=1063&format=png&auto=webp&s=94ef1c691b1025d0e6dd795e8fc1365ad022a897
To debug this I boiled it down to one prompt that scout did consistently answer correct and Maverick failed:
Prompt:
>If it takes 50 machines 5 minutes to make 5 widgets, how long would it take 100 machines to make 100 widgets?
Scout response (which is the correct answer. Keep in mind that this is "non-tricky" trick question)
>... The final answer is: $\\boxed{50}$
Maverick reponse:
>The final answer is: $\\boxed{5}$
To make sure its not an issue with the provider, I tried together, fireworks, parasail and Deepinfra on Openrouter with consistent results.
For reference, also llama 405b:
>Therefore, it would take 100 machines 50 minutes to make 100 widgets.
Noting that Maverick also [failed to impress in other benchmarks ](https://old.reddit.com/r/LocalLLaMA/comments/1jseqbs/llama_4_scout_is_not_doing_well_in_write_a/)makes me wonder whether there is an issues with the checkpoint. This evaluation should be sensitivie to pretraining, but also to RL finetuning for reasoning, as reasoning models are able to correct initial misconceptions.
Here is a prompt-by-prompt comparison.
https://preview.redd.it/bmdq7fzo16te1.png?width=4170&format=png&auto=webp&s=1ad8d1fc84eb8bf98eafe7b825a2088870497f48
Further results in the [eval folder of the repository](https://github.com/cpldcpu/MisguidedAttention/tree/main/eval)
| 2025-04-06T07:25:00 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsom5m/llama4_maverick_seems_to_perform_consistently/
|
cpldcpu
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsom5m
| false | null |
t3_1jsom5m
|
/r/LocalLLaMA/comments/1jsom5m/llama4_maverick_seems_to_perform_consistently/
| false | false | 55 |
{'enabled': False, 'images': [{'id': '0jRaPNBHNjYC2hWpicT3UC0azS4EV19kIu74M1_WoN4', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/8TxAm35XAOw6I-ky6Q-vrQzwlvyqtzxyORgJmaxuLco.jpg?width=108&crop=smart&auto=webp&s=133483b9f53abc3080d3fb49d1043238af5974d6', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/8TxAm35XAOw6I-ky6Q-vrQzwlvyqtzxyORgJmaxuLco.jpg?width=216&crop=smart&auto=webp&s=e013101f6be24114b7954fec26b49a6007650990', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/8TxAm35XAOw6I-ky6Q-vrQzwlvyqtzxyORgJmaxuLco.jpg?width=320&crop=smart&auto=webp&s=de3e6a22e0772caf1eea19763248ff2252a59826', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/8TxAm35XAOw6I-ky6Q-vrQzwlvyqtzxyORgJmaxuLco.jpg?width=640&crop=smart&auto=webp&s=96fffdc55d6363973db7d832bf4e38df09e91580', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/8TxAm35XAOw6I-ky6Q-vrQzwlvyqtzxyORgJmaxuLco.jpg?width=960&crop=smart&auto=webp&s=9658006b68eda133016702434ff7edb58cdbd5d7', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/8TxAm35XAOw6I-ky6Q-vrQzwlvyqtzxyORgJmaxuLco.jpg?width=1080&crop=smart&auto=webp&s=241b6c655469859b95f0828e8bf7cfc1542793e5', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/8TxAm35XAOw6I-ky6Q-vrQzwlvyqtzxyORgJmaxuLco.jpg?auto=webp&s=93f752786e5a9c2d9a5f055866ede60b40e03362', 'width': 1200}, 'variants': {}}]}
|
|
Orchestrator in Agentic scenario
| 1 |
I have to setup an Agentic scenario, where the orchestrator should have to dispatch the tasks based on some specific criteria, let's say deterministic, by topic. Maybe a prompt it's not enough reliable for this, so I wonder if it's a good option to make a function call, to an easy-to-maintain file (JSON), where the I can have my rule more clear.
Is it a good approach, any other ?
| 2025-04-06T07:25:38 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsomgw/orchestrator_in_agentic_scenario/
|
LuganBlan
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsomgw
| false | null |
t3_1jsomgw
|
/r/LocalLLaMA/comments/1jsomgw/orchestrator_in_agentic_scenario/
| false | false |
self
| 1 | null |
You are Llama 4.
| 1 |
[removed]
| 2025-04-06T07:28:54 |
a-third-eye
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsoo2x
| false | null |
t3_1jsoo2x
|
/r/LocalLLaMA/comments/1jsoo2x/you_are_llama_4/
| false | false | 1 |
{'enabled': True, 'images': [{'id': '7UofwqIc7qGAtSiVRMHEf7JZ0IZbCk_rX2cBBm23Lck', 'resolutions': [{'height': 46, 'url': 'https://preview.redd.it/ovdpi3rm26te1.png?width=108&crop=smart&auto=webp&s=336629fcd531b5b58229887ff3185fce4f2ef955', 'width': 108}, {'height': 92, 'url': 'https://preview.redd.it/ovdpi3rm26te1.png?width=216&crop=smart&auto=webp&s=46008f9dede91949094454f7fb42ac1254ef13ca', 'width': 216}, {'height': 137, 'url': 'https://preview.redd.it/ovdpi3rm26te1.png?width=320&crop=smart&auto=webp&s=2395c48675c3dfd9711b6caf175ee293285f4ca2', 'width': 320}, {'height': 275, 'url': 'https://preview.redd.it/ovdpi3rm26te1.png?width=640&crop=smart&auto=webp&s=965c8e6c0cb9ce7910bda81a071328f3280da2da', 'width': 640}, {'height': 412, 'url': 'https://preview.redd.it/ovdpi3rm26te1.png?width=960&crop=smart&auto=webp&s=1b0d26e425cd9ffe09cacf5e83dbb90b46b8eaa2', 'width': 960}, {'height': 464, 'url': 'https://preview.redd.it/ovdpi3rm26te1.png?width=1080&crop=smart&auto=webp&s=b7f4d78eca9bdc928bdb319ada71e5df45267405', 'width': 1080}], 'source': {'height': 884, 'url': 'https://preview.redd.it/ovdpi3rm26te1.png?auto=webp&s=3a20e1cd1757a464a1e300a447f5056d6a525796', 'width': 2056}, 'variants': {}}]}
|
||
Not the right sub but still: Is Meta a good long-term stock?
| 1 |
[removed]
| 2025-04-06T07:32:45 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsoq0u/not_the_right_sub_but_still_is_meta_a_good/
|
ultrapcb
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsoq0u
| false | null |
t3_1jsoq0u
|
/r/LocalLLaMA/comments/1jsoq0u/not_the_right_sub_but_still_is_meta_a_good/
| false | false |
self
| 1 | null |
What's your ideal mid-weight model size (20B to 33B), and why?
| 10 |
Some of my favorite models have run in this range. They seem like a good compromise between competence, speed, and memory requirements.
Contemplating this, I realized that my standards for these attributes are perhaps unusual. I have high tolerance for slow inference, frequently inferring quite happily on pure CPU (which is very slow). Also, my main for-inference GPU is an MI60 with 32GB of VRAM, which can accomodate fairly large mid-sized models with only moderate quantization.
That made me wonder what other people's standards are, and why. What are some more typical GPU VRAM sizes which can accommodate mid-sized models, and how large of a model can they handle while leaving enough memory for adequate context?
This is half idle curiosity, but also relevant to a new project I recently took up, of applying the Tulu3 post-training process to Phi-4-25B, a self-merge of Phi-4 (14B). For me 25B quantized to Q4_K_M is just about perfectly centered in my happy place, but would anyone else even use it?
| 2025-04-06T07:38:03 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsosm9/whats_your_ideal_midweight_model_size_20b_to_33b/
|
ttkciar
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsosm9
| false | null |
t3_1jsosm9
|
/r/LocalLLaMA/comments/1jsosm9/whats_your_ideal_midweight_model_size_20b_to_33b/
| false | false |
self
| 10 | null |
Best LLM/AI discord server you've found?
| 1 |
[removed]
| 2025-04-06T08:08:01 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsp7ff/best_llmai_discord_server_youve_found/
|
Warm_Iron_273
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsp7ff
| false | null |
t3_1jsp7ff
|
/r/LocalLLaMA/comments/1jsp7ff/best_llmai_discord_server_youve_found/
| false | false |
self
| 1 | null |
RL needs is a small amount of data to train multiple episodes?
| 1 |
[removed]
| 2025-04-06T08:15:02 |
https://www.reddit.com/r/LocalLLaMA/comments/1jspawl/rl_needs_is_a_small_amount_of_data_to_train/
|
External-Scratch2302
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jspawl
| false | null |
t3_1jspawl
|
/r/LocalLLaMA/comments/1jspawl/rl_needs_is_a_small_amount_of_data_to_train/
| false | false | 1 |
{'enabled': False, 'images': [{'id': '9MvjtUow7db62rL1nxV9HWnr7TDVasligixdLirI1fM', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/Wo1ECtO0R-5oaLmLcV5y6Xb3XEw3OSroTxa3z_lX_sY.jpg?width=108&crop=smart&auto=webp&s=878076cc37785f52406ed0d878ef14e665620e7d', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/Wo1ECtO0R-5oaLmLcV5y6Xb3XEw3OSroTxa3z_lX_sY.jpg?width=216&crop=smart&auto=webp&s=c3da0d655281c128bbc9531724eb231f3cb9ccbd', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/Wo1ECtO0R-5oaLmLcV5y6Xb3XEw3OSroTxa3z_lX_sY.jpg?width=320&crop=smart&auto=webp&s=189bf211158c1cbd77a5a63975f6ac1d62678902', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/Wo1ECtO0R-5oaLmLcV5y6Xb3XEw3OSroTxa3z_lX_sY.jpg?width=640&crop=smart&auto=webp&s=fd9ad1eaabab33b694f55149ca54b267703d9c17', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/Wo1ECtO0R-5oaLmLcV5y6Xb3XEw3OSroTxa3z_lX_sY.jpg?width=960&crop=smart&auto=webp&s=e3ad3d0b7fb182d182c5af4166d325c8884d1d9b', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/Wo1ECtO0R-5oaLmLcV5y6Xb3XEw3OSroTxa3z_lX_sY.jpg?width=1080&crop=smart&auto=webp&s=1919ac88bd4d9adf231b1da1790d3836bddd8536', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/Wo1ECtO0R-5oaLmLcV5y6Xb3XEw3OSroTxa3z_lX_sY.jpg?auto=webp&s=8bd4d464fec76af5684a971a5a1bbd684d262f21', 'width': 1200}, 'variants': {}}]}
|
|
Make RL Great Again
| 1 |
[removed]
| 2025-04-06T08:16:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1jspbej/make_rl_great_again/
|
External-Scratch2302
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jspbej
| false | null |
t3_1jspbej
|
/r/LocalLLaMA/comments/1jspbej/make_rl_great_again/
| false | false |
self
| 1 | null |
Two months later and after LLaMA 4's release, I'm starting to believe that supposed employee leak... Hopefully LLaMA 4's reasoning is good, because things aren't looking good for Meta.
| 450 | 2025-04-06T08:16:46 |
https://www.reddit.com/r/LocalLLaMA/comments/1jspbqk/two_months_later_and_after_llama_4s_release_im/
|
Ill-Association-8410
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jspbqk
| false | null |
t3_1jspbqk
|
/r/LocalLLaMA/comments/1jspbqk/two_months_later_and_after_llama_4s_release_im/
| false | false | 450 | null |
||
Llama 4 confusing names
| 5 |
Already started mixing up and confusing the names
| 2025-04-06T08:22:16 |
https://i.redd.it/k57tkeric6te1
|
AOHKH
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jspef9
| false | null |
t3_1jspef9
|
/r/LocalLLaMA/comments/1jspef9/llama_4_confusing_names/
| false | false |
default
| 5 |
{'enabled': False, 'images': [{'id': 'k57tkeric6te1', 'resolutions': [{'height': 44, 'url': 'https://preview.redd.it/k57tkeric6te1?width=108&crop=smart&auto=webp&s=af246232f31b0d1cd9bb8b3d629af7bd97b75591', 'width': 108}, {'height': 88, 'url': 'https://preview.redd.it/k57tkeric6te1?width=216&crop=smart&auto=webp&s=cb0446b3ca4403c9991a5317c16679baf3d1a348', 'width': 216}, {'height': 131, 'url': 'https://preview.redd.it/k57tkeric6te1?width=320&crop=smart&auto=webp&s=e69eaf84012ad5ab346c0d1bac9df9b6117d7222', 'width': 320}, {'height': 263, 'url': 'https://preview.redd.it/k57tkeric6te1?width=640&crop=smart&auto=webp&s=5ee99c6101dfa00796966e15313891fd5271ddee', 'width': 640}, {'height': 394, 'url': 'https://preview.redd.it/k57tkeric6te1?width=960&crop=smart&auto=webp&s=289af54755df91e1c3a2b8e22f5a685693bee277', 'width': 960}, {'height': 444, 'url': 'https://preview.redd.it/k57tkeric6te1?width=1080&crop=smart&auto=webp&s=3569c590b9deb5df93b3e8557f96acfa3883f784', 'width': 1080}], 'source': {'height': 496, 'url': 'https://preview.redd.it/k57tkeric6te1?auto=webp&s=c1e8dea22b8b047ac0ab5aaad923c2d34b5d3040', 'width': 1206}, 'variants': {}}]}
|
Poll: What Would It Take for You to Abandon Local AI for the Cloud?
| 0 |
Hypothetical scenario: If you were required to **permanently stop using local AI models** (like Llama) and switch *exclusively* to cloud-based alternatives, whatβs the **minimum one-time payment** youβd need to accept this change?
Consider factors like privacy, customization, offline access, and upfront hardware costs when deciding. This is just for fun β no judgment!"
**Poll Options:**
- <$10,000
- $100,000
- $100,000,000+
| 2025-04-06T08:28:44 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsphmn/poll_what_would_it_take_for_you_to_abandon_local/
|
Reasonable-Delay4740
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsphmn
| false | null |
t3_1jsphmn
|
/r/LocalLLaMA/comments/1jsphmn/poll_what_would_it_take_for_you_to_abandon_local/
| false | false |
self
| 0 | null |
LLaMa 4 (HF gated repository) access rejected in EU?
| 1 |
[removed]
| 2025-04-06T08:29:37 |
https://www.reddit.com/r/LocalLLaMA/comments/1jspi0q/llama_4_hf_gated_repository_access_rejected_in_eu/
|
m0nsky
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jspi0q
| false | null |
t3_1jspi0q
|
/r/LocalLLaMA/comments/1jspi0q/llama_4_hf_gated_repository_access_rejected_in_eu/
| false | false |
self
| 1 | null |
Has anyone used GPUStack?
| 1 |
[removed]
| 2025-04-06T08:30:39 |
https://github.com/gpustack/gpustack
|
aquarat
|
github.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jspikr
| false | null |
t3_1jspikr
|
/r/LocalLLaMA/comments/1jspikr/has_anyone_used_gpustack/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'Y6asUe8GXN4ZEOdac7fJC-F5TVPIkYX-O2jgKILsI10', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/9RiwPfYtX2uFOWhcJniMP2z_EDPUOWfu2jJIyXXPP64.jpg?width=108&crop=smart&auto=webp&s=055df8dbc953ad68df4f7f71144dd4908258a349', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/9RiwPfYtX2uFOWhcJniMP2z_EDPUOWfu2jJIyXXPP64.jpg?width=216&crop=smart&auto=webp&s=0b80f647d8c819d42a38c3c49c288b42ed0b29d4', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/9RiwPfYtX2uFOWhcJniMP2z_EDPUOWfu2jJIyXXPP64.jpg?width=320&crop=smart&auto=webp&s=5fbd2b22dcdb99d9b6361d216ffc958f940ab37a', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/9RiwPfYtX2uFOWhcJniMP2z_EDPUOWfu2jJIyXXPP64.jpg?width=640&crop=smart&auto=webp&s=40fe0fb0576f5770b822b9747f2365ec70e0e03c', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/9RiwPfYtX2uFOWhcJniMP2z_EDPUOWfu2jJIyXXPP64.jpg?width=960&crop=smart&auto=webp&s=9f899a50961a25bd1f974c8a16800df078560394', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/9RiwPfYtX2uFOWhcJniMP2z_EDPUOWfu2jJIyXXPP64.jpg?width=1080&crop=smart&auto=webp&s=8398f47d54be8eecb649ca209b558b563a4c38be', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/9RiwPfYtX2uFOWhcJniMP2z_EDPUOWfu2jJIyXXPP64.jpg?auto=webp&s=72af2b1353353f11e5db28d8f13d3a0922a30e8b', 'width': 1200}, 'variants': {}}]}
|
|
Mirrors for llama 4?
| 3 | 2025-04-06T08:31:58 |
https://www.reddit.com/r/LocalLLaMA/comments/1jspj7o/mirrors_for_llama_4/
|
k_means_clusterfuck
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jspj7o
| false | null |
t3_1jspj7o
|
/r/LocalLLaMA/comments/1jspj7o/mirrors_for_llama_4/
| false | false | 3 | null |
||
We are Open Sourcing our T-rex-mini [Roleplay] model at Saturated Labs
| 29 |
[Trex-mini](https://preview.redd.it/rl050ax7e6te1.jpg?width=1024&format=pjpg&auto=webp&s=2348eb4b2d5f02fdac8f5397ff86b4ff046bf056)
**Huggingface Link:** [Visit Here](https://huggingface.co/saturated-labs/T-Rex-mini)
Hey guys, we are open sourcing T-rex-mini model and I can say this is "the best" 8b model, it follows the instruction well and always remains in character.
**Recommend Settings/Config:**
Temperature: 1.35
min_p: 0.1
top_p: 1.0
min_p: 0.1
presence_penalty: 0.0
frequency_penalty: 0.0
repetition_penalty: 1.0
Id love to hear your feedbacks and I hope you will like it :)
**Some Backstory ( If you wanna read ):**
I am a college student I really loved to use [c.ai](http://c.ai) but overtime it really became hard to use it due to low quality response, characters will speak random things it was really frustrating, I found some alternatives but I wasn't really happy so I decided to make a research group with my friend [saturated.in](http://saturated.in) and created [loremate.saturated.in](http://loremate.saturated.in) and got really good feedbacks and many people asked us to open source it was a really hard choice as I never built anything open source, not only that I never built that people actually useπ
so I decided to open-source T-rex-mini (saturated-labs/T-Rex-mini) if the response is good we are also planning to open source other model too so please test the model and share your feedbacks :)
| 2025-04-06T08:43:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1jspow8/we_are_open_sourcing_our_trexmini_roleplay_model/
|
me_broke
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jspow8
| false | null |
t3_1jspow8
|
/r/LocalLLaMA/comments/1jspow8/we_are_open_sourcing_our_trexmini_roleplay_model/
| false | false | 29 |
{'enabled': False, 'images': [{'id': 'iOG4-s8f7_Xs9CrGPwKNp8Q9ij_1RZn_qskGpKW5XQw', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/-fNUidliPO6McDEmDz12Ut_f3RePBKDZkwQmyWVv-SA.jpg?width=108&crop=smart&auto=webp&s=7b598981bc6253e3816b87b7b680444f8b4d8c7f', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/-fNUidliPO6McDEmDz12Ut_f3RePBKDZkwQmyWVv-SA.jpg?width=216&crop=smart&auto=webp&s=92a6a3fc46662ed628fcd1553a429b0575eafb5e', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/-fNUidliPO6McDEmDz12Ut_f3RePBKDZkwQmyWVv-SA.jpg?width=320&crop=smart&auto=webp&s=464c1e1aaea4d826179894eea039a702b8e600fb', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/-fNUidliPO6McDEmDz12Ut_f3RePBKDZkwQmyWVv-SA.jpg?width=640&crop=smart&auto=webp&s=4ebeb9a5a2539f4ce181d77cabe8f99e18ceedba', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/-fNUidliPO6McDEmDz12Ut_f3RePBKDZkwQmyWVv-SA.jpg?width=960&crop=smart&auto=webp&s=955cf43799aac0babe8191659bf7884d10c769d6', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/-fNUidliPO6McDEmDz12Ut_f3RePBKDZkwQmyWVv-SA.jpg?width=1080&crop=smart&auto=webp&s=995a496e68feacfe9ec1dba0d002beb734996a1f', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/-fNUidliPO6McDEmDz12Ut_f3RePBKDZkwQmyWVv-SA.jpg?auto=webp&s=5b6c82bff1d3c7a3a23d166e60ccf0a89118a238', 'width': 1200}, 'variants': {}}]}
|
|
Has anyone used GPUStack?
| 1 |
[removed]
| 2025-04-06T08:52:41 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsptal/has_anyone_used_gpustack/
|
aquarat
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsptal
| false | null |
t3_1jsptal
|
/r/LocalLLaMA/comments/1jsptal/has_anyone_used_gpustack/
| false | false |
self
| 1 | null |
When will llama4 be released in the EU?
| 1 |
[removed]
| 2025-04-06T09:08:19 |
Wild-Masterpiece3762
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsq0xd
| false | null |
t3_1jsq0xd
|
/r/LocalLLaMA/comments/1jsq0xd/when_will_llama4_be_released_in_the_eu/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'mMrqBt0TzqoOFK3YK03DTW2Um53xV00AZw1BXPzq7L4', 'resolutions': [{'height': 182, 'url': 'https://preview.redd.it/jnmwbdiqk6te1.jpeg?width=108&crop=smart&auto=webp&s=43d229869bebd4e65be357a8249e47dad334edc4', 'width': 108}, {'height': 364, 'url': 'https://preview.redd.it/jnmwbdiqk6te1.jpeg?width=216&crop=smart&auto=webp&s=00481a809270698ff3b0cb6c4e4b17d41e653d1a', 'width': 216}, {'height': 539, 'url': 'https://preview.redd.it/jnmwbdiqk6te1.jpeg?width=320&crop=smart&auto=webp&s=8b68f1b2e5ee508f0a82a4239a7a3d0bec16b652', 'width': 320}, {'height': 1079, 'url': 'https://preview.redd.it/jnmwbdiqk6te1.jpeg?width=640&crop=smart&auto=webp&s=c5a8e5e8acdaecf2621bf28ccb989a38adf7db9d', 'width': 640}, {'height': 1619, 'url': 'https://preview.redd.it/jnmwbdiqk6te1.jpeg?width=960&crop=smart&auto=webp&s=a00500f27702a0001dd83c9c96db2ac83affe9f5', 'width': 960}, {'height': 1822, 'url': 'https://preview.redd.it/jnmwbdiqk6te1.jpeg?width=1080&crop=smart&auto=webp&s=c3cd829e0745101118420b0a9978fe116175219e', 'width': 1080}], 'source': {'height': 1822, 'url': 'https://preview.redd.it/jnmwbdiqk6te1.jpeg?auto=webp&s=2c742ee96465ffc527187c2875b0e04e60af6317', 'width': 1080}, 'variants': {}}]}
|
||
As a dual RTX 3090 owner seeing the LLaMa-4 release
| 0 | 2025-04-06T09:09:12 |
Severin_Suveren
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsq1bf
| false | null |
t3_1jsq1bf
|
/r/LocalLLaMA/comments/1jsq1bf/as_a_dual_rtx_3090_owner_seeing_the_llama4_release/
| false | false | 0 |
{'enabled': True, 'images': [{'id': 'QNYaFVgjRn63hBVTJhFy5D_RxWa4I7qz_roqkSyE1-Y', 'resolutions': [{'height': 139, 'url': 'https://preview.redd.it/vasok2mkk6te1.png?width=108&crop=smart&auto=webp&s=5b88eb691b42c05fbf9cb6ea969bc80cd14356fa', 'width': 108}, {'height': 278, 'url': 'https://preview.redd.it/vasok2mkk6te1.png?width=216&crop=smart&auto=webp&s=96d856ae5b757f2667c1692d74bbd338e026de2d', 'width': 216}, {'height': 413, 'url': 'https://preview.redd.it/vasok2mkk6te1.png?width=320&crop=smart&auto=webp&s=2ea8d2f86bc1cd6401ce3d495e758f450bce9aa8', 'width': 320}, {'height': 826, 'url': 'https://preview.redd.it/vasok2mkk6te1.png?width=640&crop=smart&auto=webp&s=2fe9346e57eaf00980a84138cced55480ca513cb', 'width': 640}, {'height': 1239, 'url': 'https://preview.redd.it/vasok2mkk6te1.png?width=960&crop=smart&auto=webp&s=19d8820492f58225ca8737acf6f9ef478192fca1', 'width': 960}, {'height': 1394, 'url': 'https://preview.redd.it/vasok2mkk6te1.png?width=1080&crop=smart&auto=webp&s=eccf6ea63995972006f7939428f80db5abd52da6', 'width': 1080}], 'source': {'height': 1401, 'url': 'https://preview.redd.it/vasok2mkk6te1.png?auto=webp&s=1b080583ca893a7ff554b8271e086685b75df2ab', 'width': 1085}, 'variants': {}}]}
|
|||
Smaller Gemma3 QAT versions: 12B in < 8GB and 27B in <16GB !
| 253 |
I was a bit frustrated by the release of Gemma3 QAT (quantized-aware training). These models are performing insanely well for quantized models, but despite being advertised as "q4\_0" quants, they were bigger than some 5-bit quants out there, and critically, they were above the 16GB and 8GB thresholds for the 27B and 12B models respectively, which makes them harder to run fully offloaded to some consumer GPUS.
I quickly found out that the reason for this significant size increease compared to normal q4\_0 quants was the unquantized, half precision token embeddings table, wheras, by llama.cpp standards, this table should be quantized to Q6\_K type.
So I did some "brain surgery" and swapped out the embeddings table from those QAT models with the one taken from an imatrix-quantized model by [bartowski](https://huggingface.co/bartowski). The end product is a model that is performing almost exactly like the "full" QAT model by google, but significantly smaller. I ran some perplexity tests, and the results were consistently within margin of error.
You can find the weights (and the script I used to perform the surgery) here:
[https://huggingface.co/stduhpf/google-gemma-3-27b-it-qat-q4\_0-gguf-small](https://huggingface.co/stduhpf/google-gemma-3-27b-it-qat-q4_0-gguf-small)
[https://huggingface.co/stduhpf/google-gemma-3-12b-it-qat-q4\_0-gguf-small](https://huggingface.co/stduhpf/google-gemma-3-12b-it-qat-q4_0-gguf-small)
With these I can run Gemma3 12b qat on a 8GB GPU with 2.5k context window without any other optimisation, and by enabling flash attention and q8 kv cache, it can go up to 4k ctx.
| 2025-04-06T09:10:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsq1so/smaller_gemma3_qat_versions_12b_in_8gb_and_27b_in/
|
stduhpf
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsq1so
| false | null |
t3_1jsq1so
|
/r/LocalLLaMA/comments/1jsq1so/smaller_gemma3_qat_versions_12b_in_8gb_and_27b_in/
| false | false |
self
| 253 |
{'enabled': False, 'images': [{'id': 'oNZUrCbKtO_M_U66k1LYvlCFzmkPOiQnoPXqwNClTvg', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/clkgEJmdj-4gAgRhhH1bSXUy-XypdLf4tBI0ZR-lfvM.jpg?width=108&crop=smart&auto=webp&s=363cd79195ec3e35bad07940d2fcd0a254d1b753', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/clkgEJmdj-4gAgRhhH1bSXUy-XypdLf4tBI0ZR-lfvM.jpg?width=216&crop=smart&auto=webp&s=142c6a940bbae65b1b74e6e545b76e93b955ea5a', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/clkgEJmdj-4gAgRhhH1bSXUy-XypdLf4tBI0ZR-lfvM.jpg?width=320&crop=smart&auto=webp&s=1bfdfb49f94f5566f34016e5e4548c9cf41a7a26', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/clkgEJmdj-4gAgRhhH1bSXUy-XypdLf4tBI0ZR-lfvM.jpg?width=640&crop=smart&auto=webp&s=b82905ff0d883b36d72d0366d40149d075ea2451', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/clkgEJmdj-4gAgRhhH1bSXUy-XypdLf4tBI0ZR-lfvM.jpg?width=960&crop=smart&auto=webp&s=4ad1975dc90014276aec6aefb8230ad50acfc6b5', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/clkgEJmdj-4gAgRhhH1bSXUy-XypdLf4tBI0ZR-lfvM.jpg?width=1080&crop=smart&auto=webp&s=e2303ed0d223329bdf39cd8a8f39cbf3795407af', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/clkgEJmdj-4gAgRhhH1bSXUy-XypdLf4tBI0ZR-lfvM.jpg?auto=webp&s=0571810d24cdb6d2f44400441a52c84d1e25fa59', 'width': 1200}, 'variants': {}}]}
|
My knowledge extraction benchmark has both Llama 4 on the podium
| 1 | 2025-04-06T09:11:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsq2ko/my_knowledge_extraction_benchmark_has_both_llama/
|
olddoglearnsnewtrick
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsq2ko
| false | null |
t3_1jsq2ko
|
/r/LocalLLaMA/comments/1jsq2ko/my_knowledge_extraction_benchmark_has_both_llama/
| false | false | 1 | null |
||
Its about time you guys stop trying to run models at home
| 1 |
[removed]
| 2025-04-06T09:13:16 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsq3ez/its_about_time_you_guys_stop_trying_to_run_models/
|
Defiant-Mood6717
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsq3ez
| false | null |
t3_1jsq3ez
|
/r/LocalLLaMA/comments/1jsq3ez/its_about_time_you_guys_stop_trying_to_run_models/
| false | false |
self
| 1 | null |
Discussion - It's about time you guys stop trying to run models at home
| 1 |
[removed]
| 2025-04-06T09:15:33 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsq4i0/discussion_its_about_time_you_guys_stop_trying_to/
|
Defiant-Mood6717
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsq4i0
| false | null |
t3_1jsq4i0
|
/r/LocalLLaMA/comments/1jsq4i0/discussion_its_about_time_you_guys_stop_trying_to/
| false | false |
self
| 1 | null |
What rule did I violate with this post? These moderators are terrible and don't allow free speech
| 1 |
[removed]
| 2025-04-06T09:18:07 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsq5u7/what_rule_did_i_violate_with_this_post_these/
|
Defiant-Mood6717
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsq5u7
| false | null |
t3_1jsq5u7
|
/r/LocalLLaMA/comments/1jsq5u7/what_rule_did_i_violate_with_this_post_these/
| false | false | 1 | null |
|
Coder coder coder coder
| 0 |
I love all these open sources/weights models from companies or individuals but I feel like what is really missing is a specialized « middle size » coder => built for coding coding coding coding and exploding main closed models.
Sometimes I dream some experts in the locallama community (or hf team) saying: ok letβs not wait for others to do the job, letβs create our open high quality coding dataset, and guiding noob people like me how to help/participate to that big project. Even if that makes no sense, lemme dream :)
| 2025-04-06T09:22:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsq81g/coder_coder_coder_coder/
|
Leflakk
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsq81g
| false | null |
t3_1jsq81g
|
/r/LocalLLaMA/comments/1jsq81g/coder_coder_coder_coder/
| false | false |
self
| 0 | null |
Is there anyone tried Running Deepseek r1 on cpu ram only?
| 1 |
[removed]
| 2025-04-06T09:27:46 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsqanx/is_there_anyone_tried_running_deepseek_r1_on_cpu/
|
FamousAdvertising550
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsqanx
| false | null |
t3_1jsqanx
|
/r/LocalLLaMA/comments/1jsqanx/is_there_anyone_tried_running_deepseek_r1_on_cpu/
| false | false |
self
| 1 | null |
Explore GitHub repos, PDF docs better through mindmaps, RAG based search
| 1 |
[removed]
| 2025-04-06T09:34:46 |
https://v.redd.it/hirh4ozfo6te1
|
Remarkable-Pair-4424
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsqe25
| false |
{'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/hirh4ozfo6te1/DASHPlaylist.mpd?a=1746524105%2CYjE2NTYwZDMzNGU1OGMyMDgyOTI4MTQwNDIxMjIxZTlhMjFhYmI4YmRjMGY4NTIwZTA0YmUxOTMzODMwNDA2ZA%3D%3D&v=1&f=sd', 'duration': 71, 'fallback_url': 'https://v.redd.it/hirh4ozfo6te1/DASH_1080.mp4?source=fallback', 'has_audio': False, 'height': 1080, 'hls_url': 'https://v.redd.it/hirh4ozfo6te1/HLSPlaylist.m3u8?a=1746524105%2CYWUyNzY5N2ZkZTgzMDI2NDc1MDljZjZkZGQzNzBjNjI1MDFhODgyOGRhZGZhZWIyNDQ5YzY3M2I4YzRkMjg3Ng%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/hirh4ozfo6te1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1920}}
|
t3_1jsqe25
|
/r/LocalLLaMA/comments/1jsqe25/explore_github_repos_pdf_docs_better_through/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'cjZ4cjJvemZvNnRlMeXrz-mtEUx8WS9k2S86S-lwtDjI2ALmjxeB6ln-mIhZ', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/cjZ4cjJvemZvNnRlMeXrz-mtEUx8WS9k2S86S-lwtDjI2ALmjxeB6ln-mIhZ.png?width=108&crop=smart&format=pjpg&auto=webp&s=6fea440accb636e83206cc113ca161397477fbd2', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/cjZ4cjJvemZvNnRlMeXrz-mtEUx8WS9k2S86S-lwtDjI2ALmjxeB6ln-mIhZ.png?width=216&crop=smart&format=pjpg&auto=webp&s=d0d4b075398f815754908aabe1caa6d5ec1dd6ae', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/cjZ4cjJvemZvNnRlMeXrz-mtEUx8WS9k2S86S-lwtDjI2ALmjxeB6ln-mIhZ.png?width=320&crop=smart&format=pjpg&auto=webp&s=642e32586de50c79ca6da8d681bbb0b99f711f3e', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/cjZ4cjJvemZvNnRlMeXrz-mtEUx8WS9k2S86S-lwtDjI2ALmjxeB6ln-mIhZ.png?width=640&crop=smart&format=pjpg&auto=webp&s=9efcfc9b6b25d36ae9b9bc62defe95eddba18946', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/cjZ4cjJvemZvNnRlMeXrz-mtEUx8WS9k2S86S-lwtDjI2ALmjxeB6ln-mIhZ.png?width=960&crop=smart&format=pjpg&auto=webp&s=9b9feb3180df75ba2e955dec2b202ff2b93fb283', 'width': 960}, {'height': 607, 'url': 'https://external-preview.redd.it/cjZ4cjJvemZvNnRlMeXrz-mtEUx8WS9k2S86S-lwtDjI2ALmjxeB6ln-mIhZ.png?width=1080&crop=smart&format=pjpg&auto=webp&s=9aacdce6bfeb3d0c08032cd180d0d915de0f5609', 'width': 1080}], 'source': {'height': 1080, 'url': 'https://external-preview.redd.it/cjZ4cjJvemZvNnRlMeXrz-mtEUx8WS9k2S86S-lwtDjI2ALmjxeB6ln-mIhZ.png?format=pjpg&auto=webp&s=f9e2cbf0728694ecff45e3c60cba570495c312d6', 'width': 1920}, 'variants': {}}]}
|
|
Big moe models => cpu/mac inference?
| 2 |
With the advent of all these big moe, with a resonnable budget we're kind of forced from multi gpu inference to cpu or mac inference. How do you feel about that? Do you think it will be a long lasting trend?
First time I saw a big moe as such was the very first grok iirc, but I feel we'll see much more of these, which completely changes the hardware paradigm for us in localllama.
Another take would be to use these huge models as foundational models and wait for them to be distilled in others smaller models. May be the times of good crazy fine-tunes is back?!
I can't fathom the sort of gpu node needed to finetune these.. you already need a beefy one just to generate a synthetic dataset with them π
| 2025-04-06T09:37:07 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsqf68/big_moe_models_cpumac_inference/
|
No_Afternoon_4260
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsqf68
| false | null |
t3_1jsqf68
|
/r/LocalLLaMA/comments/1jsqf68/big_moe_models_cpumac_inference/
| false | false |
self
| 2 | null |
LLaMa 4 completely flops at my linguistic usecase
| 26 |
Just tried Maverick on a task: given a sentence in a foreign language, explain each word in it by giving a contextual translation.
It can't even format the output correctly (I guide LLMs to the correct formatting with prompting and also provide examples; much smaller models are able to do that).
| 2025-04-06T09:37:36 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsqfes/llama_4_completely_flops_at_my_linguistic_usecase/
|
schattig_eenhoorntje
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsqfes
| false | null |
t3_1jsqfes
|
/r/LocalLLaMA/comments/1jsqfes/llama_4_completely_flops_at_my_linguistic_usecase/
| false | false |
self
| 26 | null |
Quick review of EXAONE Deep 32B
| 13 |
I stumbled upon this model on Ollama today, and it seems to be the only 32B reasoning model that uses RL other than QwQ.
\*QwQ passed all the following tests; see [this post](https://www.reddit.com/r/LocalLLaMA/comments/1js0zmd/quick_comparison_of_qwq_and_openthinker2_32b/) for more information. I will only post EXAONE's results here.
\---
Candle test:
>Failed [https://imgur.com/a/5Vslve4](https://imgur.com/a/5Vslve4)
5 reasoning questions:
>3 passed, 2 failed [https://imgur.com/a/4neDoea](https://imgur.com/a/4neDoea)
\---
Private tests:
Coding question: One question about what caused the issue, plus 1,200 lines of C++ code.
>Passed, however, during multi-shot testing, it has a 50% chance of failing.
Restructuring a financial spreadsheet.
>Passed.
\---
Conclusion:
Even though LG said they also used RL in their paper, this model is still noticeably weaker than QwQ.
Additionally, this model suffers from the worst "overthinking" issue I have ever seen. For example, it wrote a 3573-word essay to answer "Tell me a random fun fact about the Roman Empire." Although it never fell into a loop, it thinks longer than any local reasoning model I have ever tested, and it is highly indecisive during the thinking process.
\---
Settings I used:Β [https://imgur.com/a/7ZBQ6SX](https://imgur.com/a/7ZBQ6SX)
gguf:
[https://huggingface.co/bartowski/LGAI-EXAONE\_EXAONE-Deep-32B-GGUF/blob/main/LGAI-EXAONE\_EXAONE-Deep-32B-IQ4\_XS.gguf](https://huggingface.co/bartowski/LGAI-EXAONE_EXAONE-Deep-32B-GGUF/blob/main/LGAI-EXAONE_EXAONE-Deep-32B-IQ4_XS.gguf)
backend: ollama
source of public questions:
[https://www.reddit.com/r/LocalLLaMA/comments/1i65599/r1\_32b\_is\_be\_worse\_than\_qwq\_32b\_tests\_included/](https://www.reddit.com/r/LocalLLaMA/comments/1i65599/r1_32b_is_be_worse_than_qwq_32b_tests_included/)
[https://www.reddit.com/r/LocalLLaMA/comments/1jpr1nk/the\_candle\_test\_most\_llms\_fail\_to\_generalise\_at/](https://www.reddit.com/r/LocalLLaMA/comments/1jpr1nk/the_candle_test_most_llms_fail_to_generalise_at/)
| 2025-04-06T09:49:36 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsql79/quick_review_of_exaone_deep_32b/
|
AaronFeng47
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsql79
| false | null |
t3_1jsql79
|
/r/LocalLLaMA/comments/1jsql79/quick_review_of_exaone_deep_32b/
| false | false |
self
| 13 |
{'enabled': False, 'images': [{'id': '-tdanGdoX7QTPv8UzMpF1FMJHfU3KJEVAPqWHZSW4-E', 'resolutions': [{'height': 129, 'url': 'https://external-preview.redd.it/3-lVEXLSvzwlSPew2thrK4GGrDKy7Za5lxs3ZQVBS70.jpg?width=108&crop=smart&auto=webp&s=645e93e6e6df8f04d5b1c7a7887d742480583aa9', 'width': 108}, {'height': 259, 'url': 'https://external-preview.redd.it/3-lVEXLSvzwlSPew2thrK4GGrDKy7Za5lxs3ZQVBS70.jpg?width=216&crop=smart&auto=webp&s=104c97eb98455bd2a18ef52e841a8b3f82514f96', 'width': 216}, {'height': 384, 'url': 'https://external-preview.redd.it/3-lVEXLSvzwlSPew2thrK4GGrDKy7Za5lxs3ZQVBS70.jpg?width=320&crop=smart&auto=webp&s=46e284d2a7f0fba78d0e54b7ae30f152529140ab', 'width': 320}, {'height': 769, 'url': 'https://external-preview.redd.it/3-lVEXLSvzwlSPew2thrK4GGrDKy7Za5lxs3ZQVBS70.jpg?width=640&crop=smart&auto=webp&s=a4cd58f114586c06934cf6da941b92063c0f1145', 'width': 640}, {'height': 1154, 'url': 'https://external-preview.redd.it/3-lVEXLSvzwlSPew2thrK4GGrDKy7Za5lxs3ZQVBS70.jpg?width=960&crop=smart&auto=webp&s=78c3e615f26ba3dd82941c258b348570addc03a9', 'width': 960}], 'source': {'height': 1257, 'url': 'https://external-preview.redd.it/3-lVEXLSvzwlSPew2thrK4GGrDKy7Za5lxs3ZQVBS70.jpg?auto=webp&s=680471a73efa51961683cd8e4c9a1ec6be8986cd', 'width': 1045}, 'variants': {}}]}
|
Any ideas why they decided to release Llama 4 on Saturday instead of Monday?
| 147 | 2025-04-06T10:02:56 |
nobilix
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsqs2x
| false | null |
t3_1jsqs2x
|
/r/LocalLLaMA/comments/1jsqs2x/any_ideas_why_they_decided_to_release_llama_4_on/
| false | false | 147 |
{'enabled': True, 'images': [{'id': '6Qqinjfo9YTRxP-fzYxOy8YdEoHF2bogFmzXXQzR6AM', 'resolutions': [{'height': 207, 'url': 'https://preview.redd.it/yfyvqx7hu6te1.jpeg?width=108&crop=smart&auto=webp&s=973b52dbdfa8874dc5cae04dcf9a968365d45b1d', 'width': 108}, {'height': 415, 'url': 'https://preview.redd.it/yfyvqx7hu6te1.jpeg?width=216&crop=smart&auto=webp&s=604a9cded00abad91b6613102f273b776a56f3e9', 'width': 216}, {'height': 616, 'url': 'https://preview.redd.it/yfyvqx7hu6te1.jpeg?width=320&crop=smart&auto=webp&s=25a11a3501486b0dfa059e2de5fcbbb62ea3cdcc', 'width': 320}, {'height': 1232, 'url': 'https://preview.redd.it/yfyvqx7hu6te1.jpeg?width=640&crop=smart&auto=webp&s=dc0e570396a07e48a96887834fce2fa520710646', 'width': 640}, {'height': 1848, 'url': 'https://preview.redd.it/yfyvqx7hu6te1.jpeg?width=960&crop=smart&auto=webp&s=c0701f024c84e92bfa8d66c441b70663b451d43b', 'width': 960}, {'height': 2079, 'url': 'https://preview.redd.it/yfyvqx7hu6te1.jpeg?width=1080&crop=smart&auto=webp&s=66e0cc06abb92e2a5fe06bed12fe6b8c9e827500', 'width': 1080}], 'source': {'height': 2484, 'url': 'https://preview.redd.it/yfyvqx7hu6te1.jpeg?auto=webp&s=a8ed10f747326251c1c79313c0cac8af0637a6f5', 'width': 1290}, 'variants': {}}]}
|
|||
i wanna create a prompt engineer similar to the one in chatgpt
| 1 |
[removed]
| 2025-04-06T10:11:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsqw6h/i_wanna_create_a_prompt_engineer_similar_to_the/
|
I_am_a_menace
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsqw6h
| false | null |
t3_1jsqw6h
|
/r/LocalLLaMA/comments/1jsqw6h/i_wanna_create_a_prompt_engineer_similar_to_the/
| false | false |
self
| 1 | null |
Anyone else seeing Quasar Alpha on OpenRouter error out around 90k tokens?
| 1 |
[removed]
| 2025-04-06T10:12:14 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsqwi3/anyone_else_seeing_quasar_alpha_on_openrouter/
|
EinsteinOnRedbull
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsqwi3
| false | null |
t3_1jsqwi3
|
/r/LocalLLaMA/comments/1jsqwi3/anyone_else_seeing_quasar_alpha_on_openrouter/
| false | false |
self
| 1 | null |
Meta.ai doesn't know what Llama 4 model it is
| 0 |
With the news of Llama 4 being released I wanted to try it out so I went to meta.ai.
I asked it what llm it is and it said 4. I asked it what models and variants exist with 4 and it gave me the official list.
Then I asked which are you and you can see for yourselves what it said.
I'm not impressed.
| 2025-04-06T10:20:00 |
laffingbuddhas
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsr08w
| false | null |
t3_1jsr08w
|
/r/LocalLLaMA/comments/1jsr08w/metaai_doesnt_know_what_llama_4_model_it_is/
| false | false | 0 |
{'enabled': True, 'images': [{'id': 'srVUPSRA8zVNoJ94nHgFusRQN_XhB_wtbUsIkzzlk5Q', 'resolutions': [{'height': 79, 'url': 'https://preview.redd.it/8xm1idxix6te1.png?width=108&crop=smart&auto=webp&s=2d08c21fa51faa0a88280141cb59d75d2c4172a6', 'width': 108}, {'height': 158, 'url': 'https://preview.redd.it/8xm1idxix6te1.png?width=216&crop=smart&auto=webp&s=c6979090ee711016c870034e0700ba0e32137dd6', 'width': 216}, {'height': 234, 'url': 'https://preview.redd.it/8xm1idxix6te1.png?width=320&crop=smart&auto=webp&s=4cb58d6d3bb1d9f06dd6b2ac9bb8af1a238dfdab', 'width': 320}, {'height': 468, 'url': 'https://preview.redd.it/8xm1idxix6te1.png?width=640&crop=smart&auto=webp&s=26be38c621e352fcaa31fc64588153adb09a2f31', 'width': 640}, {'height': 703, 'url': 'https://preview.redd.it/8xm1idxix6te1.png?width=960&crop=smart&auto=webp&s=487166ec03ce46385f3fd1b9835597259c132703', 'width': 960}, {'height': 791, 'url': 'https://preview.redd.it/8xm1idxix6te1.png?width=1080&crop=smart&auto=webp&s=557c7a9045fe360bf1a51495c84a0ef1f667a49b', 'width': 1080}], 'source': {'height': 791, 'url': 'https://preview.redd.it/8xm1idxix6te1.png?auto=webp&s=3f7d6fc7b638dda84d6b40e270513891702a57de', 'width': 1080}, 'variants': {}}]}
|
||
Aider with QwQ + Qwen coder
| 7 |
I am struggling to make these models to work correctly with aider. Almost always get edit errors and never really get decent results. Can anyone that got it to work correctly say what I am doing wrong here? I downloaded the models and I am running them locally with llama-swap. here is the aider config file:
- name: "openai/qwq-32b"
edit_format: diff
extra_params:
max_tokens: 16384
top_p: 0.95
top_k: 40
presence_penalty: 0.1
repetition_penalty: 1
num_ctx: 16384
use_temperature: 0.6
weak_model_name: "openai/qwen25-coder"
editor_model_name: "openai/qwen25-coder"
reasoning_tag: think
- name: "openai/qwen25-coder"
edit_format: diff
extra_params:
max_tokens: 16000
top_p: 0.8
top_k: 20
repetition_penalty: 1.05
use_temperature: 0.7
reasoning_tag: null
editor_model_name: "openai/qwen25-coder"
editor_edit_format: editor-diff
I have tried starting aider with many different options:
aider --architect --model openai/qwq-32b --editor-model openai/qwen25-coder
Appreciate any ideas. Thanks.
| 2025-04-06T10:24:51 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsr2lw/aider_with_qwq_qwen_coder/
|
arivar
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsr2lw
| false | null |
t3_1jsr2lw
|
/r/LocalLLaMA/comments/1jsr2lw/aider_with_qwq_qwen_coder/
| false | false |
self
| 7 | null |
What are your thoughts about the Llama 4 models?
| 71 |
Its clear from Marks announcement theyre still training their bigger models. Likely they are going to gather feedback on these two and release improvements on the larger models and enhance these for their usual .1-.3 series once they realize the models are not performing up to par. With Gemini 2.5 and Claude 3.7 and the o3 series, the bar is much higher than it was for llama3. With that said, with skilled fine tuning, they might turn out to be very useful. If they really want to win, they should go full open source and let the community enhance llama and then train llama5 on those enhancements.
| 2025-04-06T10:36:50 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsr8ie/what_are_your_thoughts_about_the_llama_4_models/
|
internal-pagal
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsr8ie
| false | null |
t3_1jsr8ie
|
/r/LocalLLaMA/comments/1jsr8ie/what_are_your_thoughts_about_the_llama_4_models/
| false | false |
self
| 71 | null |
Llama 4 System Prompt Has Some Interesting Details:
| 1 |
[removed]
| 2025-04-06T11:06:57 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsroaa/llama_4_system_prompt_has_some_interesting_details/
|
BluebirdFit3951
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsroaa
| false | null |
t3_1jsroaa
|
/r/LocalLLaMA/comments/1jsroaa/llama_4_system_prompt_has_some_interesting_details/
| false | false |
self
| 1 | null |
109b vs 24b ?? What's this benchmark?
| 225 |
Like llama 4 scout is 109b parameters and they compared with 24 and 27b parameters (I'm talking about total parameters size )
| 2025-04-06T11:27:05 |
Independent-Wind4462
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsrz5v
| false | null |
t3_1jsrz5v
|
/r/LocalLLaMA/comments/1jsrz5v/109b_vs_24b_whats_this_benchmark/
| false | false | 225 |
{'enabled': True, 'images': [{'id': '2MfZWN7t-sMMxFNipGceSM15wqwZKEq3nx8eNG7itaI', 'resolutions': [{'height': 108, 'url': 'https://preview.redd.it/igg46skh97te1.png?width=108&crop=smart&auto=webp&s=83e601814a0435bdc3f38d226330574f2e754eef', 'width': 108}, {'height': 216, 'url': 'https://preview.redd.it/igg46skh97te1.png?width=216&crop=smart&auto=webp&s=7d8d07b3078147f6289dfe008b49db8f53cd4cbb', 'width': 216}, {'height': 321, 'url': 'https://preview.redd.it/igg46skh97te1.png?width=320&crop=smart&auto=webp&s=7343c7ffb963cd6e2282ac0846bd706d9c5608b4', 'width': 320}, {'height': 642, 'url': 'https://preview.redd.it/igg46skh97te1.png?width=640&crop=smart&auto=webp&s=ac1913f8538347a323b43f755d80a1b4bee7dcc0', 'width': 640}, {'height': 963, 'url': 'https://preview.redd.it/igg46skh97te1.png?width=960&crop=smart&auto=webp&s=1bdd93e7ab6dcad276419d15d4766474698dedd7', 'width': 960}, {'height': 1084, 'url': 'https://preview.redd.it/igg46skh97te1.png?width=1080&crop=smart&auto=webp&s=e4aad0a2e280f4a59154882fa6f769c840f61760', 'width': 1080}], 'source': {'height': 1084, 'url': 'https://preview.redd.it/igg46skh97te1.png?auto=webp&s=92d7f1cc33e589d1f20af887bf80a61b366ec805', 'width': 1080}, 'variants': {}}]}
|
||
A context window of 10M tokens is outrageous.
| 1 |
[removed]
| 2025-04-06T11:53:00 |
Sweaty-Scene5621
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jssdvn
| false | null |
t3_1jssdvn
|
/r/LocalLLaMA/comments/1jssdvn/a_context_window_of_10m_tokens_is_outrageous/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'C-Ojqc8m0Nj1vH0VV5v4S8aqcY7dOjTPMBfUl-VzdXY', 'resolutions': [{'height': 151, 'url': 'https://preview.redd.it/2asb2154e7te1.jpeg?width=108&crop=smart&auto=webp&s=5fa28db6e6ee040ea3909bee27371abe8f36f42d', 'width': 108}, {'height': 302, 'url': 'https://preview.redd.it/2asb2154e7te1.jpeg?width=216&crop=smart&auto=webp&s=8b8f7f1905bdd00cb9bf0d7f275277048ac2dd41', 'width': 216}, {'height': 447, 'url': 'https://preview.redd.it/2asb2154e7te1.jpeg?width=320&crop=smart&auto=webp&s=660e523958890a2f3287dc5e0cecbcd08979e93b', 'width': 320}, {'height': 895, 'url': 'https://preview.redd.it/2asb2154e7te1.jpeg?width=640&crop=smart&auto=webp&s=388a81028708faebe7bf8566928d34ca106e08b8', 'width': 640}], 'source': {'height': 1007, 'url': 'https://preview.redd.it/2asb2154e7te1.jpeg?auto=webp&s=1e48e3fe1d75116123ff3a6e183e63cf2b29f282', 'width': 720}, 'variants': {}}]}
|
||
Models for MacBook M4 Pro + How to fine-tune?
| 0 |
Hi so I recently bought the new MacBook M4 Pro with 16GB RAM, 10 core GPU and 512 SSD. I do know that the maximum I can run is 7B models. But would like your suggestions on which good models to run.
1. The project that I can aiming for is give the model some my dairy pdfs for each friend and it summarize and answer me things about them I wrote in the diary.
2. The another project is very similar but it will be based on the WhatsApp messages of each friend and family and simply respond to them.
I need suggestions for which model (censored/uncensored but not NSFW ones )to run for my first time. I know the basics of Generative AI (maximum I learnt is the Mistral 7B paper and its MoE but unable to do practicals due to many issues)
| 2025-04-06T11:54:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1jssenb/models_for_macbook_m4_pro_how_to_finetune/
|
FastCommission2913
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jssenb
| false | null |
t3_1jssenb
|
/r/LocalLLaMA/comments/1jssenb/models_for_macbook_m4_pro_how_to_finetune/
| false | false |
self
| 0 | null |
How to connect local LLM/SLM to use local datasets (Physics research papers) via MCP (Model Context Protocol)?
| 1 |
[removed]
| 2025-04-06T11:57:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1jssgj0/how_to_connect_local_llmslm_to_use_local_datasets/
|
zoxtech
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jssgj0
| false | null |
t3_1jssgj0
|
/r/LocalLLaMA/comments/1jssgj0/how_to_connect_local_llmslm_to_use_local_datasets/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'HIWMVljlXnX_Ej7OH_UZMTrZmT5iqSFH5SpJmiFg2GE', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/YrruCtJ3dSQtTg5PBMDV4DVsO_nt4gqVEkfOh1yYiCQ.jpg?width=108&crop=smart&auto=webp&s=c9f230616cc003c225a9899da5851f13c6319038', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/YrruCtJ3dSQtTg5PBMDV4DVsO_nt4gqVEkfOh1yYiCQ.jpg?width=216&crop=smart&auto=webp&s=73527270984140acded7c0c2951f6f333c7167ea', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/YrruCtJ3dSQtTg5PBMDV4DVsO_nt4gqVEkfOh1yYiCQ.jpg?width=320&crop=smart&auto=webp&s=4ba3758424723b366af0810e4d8044acb64fafbc', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/YrruCtJ3dSQtTg5PBMDV4DVsO_nt4gqVEkfOh1yYiCQ.jpg?width=640&crop=smart&auto=webp&s=c50a795025611cd3f253d1af99920c6a8046c267', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/YrruCtJ3dSQtTg5PBMDV4DVsO_nt4gqVEkfOh1yYiCQ.jpg?width=960&crop=smart&auto=webp&s=5c2aa6a0ee52a6fa2053dcf394dc704a62824182', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/YrruCtJ3dSQtTg5PBMDV4DVsO_nt4gqVEkfOh1yYiCQ.jpg?width=1080&crop=smart&auto=webp&s=7ea0ff5d7cc7589630cbcbf52fd51a00d35801ab', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/YrruCtJ3dSQtTg5PBMDV4DVsO_nt4gqVEkfOh1yYiCQ.jpg?auto=webp&s=74cbbb3c4dcb8362ca423fcd10f2f97095828cb9', 'width': 1200}, 'variants': {}}]}
|
"snugly fits in a h100, quantized 4 bit"
| 1,313 | 2025-04-06T11:59:08 |
LoSboccacc
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsshhe
| false | null |
t3_1jsshhe
|
/r/LocalLLaMA/comments/1jsshhe/snugly_fits_in_a_h100_quantized_4_bit/
| false | false | 1,313 |
{'enabled': True, 'images': [{'id': 'O_4RfCpft2_pnUyZ_q0xRZsVeEvNmUwdQDISguFB5eM', 'resolutions': [{'height': 119, 'url': 'https://preview.redd.it/g2mj9lg4f7te1.png?width=108&crop=smart&auto=webp&s=51926e2d8ef9e10a58295284e8644888ec4e2517', 'width': 108}, {'height': 239, 'url': 'https://preview.redd.it/g2mj9lg4f7te1.png?width=216&crop=smart&auto=webp&s=483b25fd827ddc9a0b80e84319c7a37fcde6d2ac', 'width': 216}, {'height': 355, 'url': 'https://preview.redd.it/g2mj9lg4f7te1.png?width=320&crop=smart&auto=webp&s=0aee8c1a42548cd1115d28de9329c9ac04f96ff5', 'width': 320}, {'height': 710, 'url': 'https://preview.redd.it/g2mj9lg4f7te1.png?width=640&crop=smart&auto=webp&s=4b3f27828a4984437b27e38b91aa497b1074ed5d', 'width': 640}, {'height': 1065, 'url': 'https://preview.redd.it/g2mj9lg4f7te1.png?width=960&crop=smart&auto=webp&s=303fbb9c3ee2c36422c3e36c50aef9a491297c45', 'width': 960}], 'source': {'height': 1137, 'url': 'https://preview.redd.it/g2mj9lg4f7te1.png?auto=webp&s=5d94801271c7a71916ddc4ea7bad7bbc0e07a2e7', 'width': 1024}, 'variants': {}}]}
|
|||
Llama4 is horrible, but when have llama been good?
| 1 |
[removed]
| 2025-04-06T12:03:30 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsskan/llama4_is_horrible_but_when_have_llama_been_good/
|
sunomonodekani
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsskan
| false | null |
t3_1jsskan
|
/r/LocalLLaMA/comments/1jsskan/llama4_is_horrible_but_when_have_llama_been_good/
| false | false |
self
| 1 | null |
Why not 16x Nvidia Tesla K80?
| 1 |
Ignore power consumption for a second. Lets say i got a motherboard with 4 of x16 pcie gen3 lanes, why couldn't I just fill it up with Nvidia Tesla K80s and run huge LLMs, they are dual gpu cards, 12gb ddr5, 4.1 TFLOPS fp16, each. 4 Cards of thoes would theoreticly be 96gb, 1924.8gb/s bandwidth, 65.6 tops. Lets go even further and say I got an enterprise motherboard, do some pcie bifuscation and now have 16 cards, x8 lanes (i dont know how doable that is). thats theoreticly 384gb total vram, 7700gb/s bandwidth, 66 tops. Assuming power is free, would this be such a bad idea, when the cards are so cheap?
| 2025-04-06T12:07:14 |
https://www.reddit.com/r/LocalLLaMA/comments/1jssmlp/why_not_16x_nvidia_tesla_k80/
|
NoBlame4You
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jssmlp
| false | null |
t3_1jssmlp
|
/r/LocalLLaMA/comments/1jssmlp/why_not_16x_nvidia_tesla_k80/
| false | false |
self
| 1 | null |
Dark Model
| 0 |
Anyone else think it's painfully obvious that the elites have a model that is fully trained on all of human kinds private data at this point?
I would bet dollars to donuts that executives at OpenAI, Deepseek, Anthropic, etc, have models available to them where they can ask any question about any specific person and get an accurate answer in seconds.
Moreover, everyone is concerned about fake news and deepfake images generated by people online, but no concern for the elites being the only ones with access to it since YOU KNOW Altman has a private ChatGPT app that can generate a photo realistic picture of anyone (real or not) doing ANYTHING
Sick and disgusting imo to allow people in power to stay in power by making them the only ones able to control this technology.
| 2025-04-06T12:40:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1jst7m5/dark_model/
|
butthole_nipple
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jst7m5
| false | null |
t3_1jst7m5
|
/r/LocalLLaMA/comments/1jst7m5/dark_model/
| false | false |
self
| 0 | null |
Specs for Llama 4 Behemot (2T)
| 0 |
Was wondering what kind of rig would Behemot require to be "summoned", quantized and unquantized?
| 2025-04-06T12:42:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1jst8jd/specs_for_llama_4_behemot_2t/
|
Neptun0
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jst8jd
| false | null |
t3_1jst8jd
|
/r/LocalLLaMA/comments/1jst8jd/specs_for_llama_4_behemot_2t/
| false | false |
self
| 0 | null |
Llama 4 is here
| 0 |
Llama 4: https://www.llama.com/llama4/
| 2025-04-06T12:44:33 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsta8y/llama_4_is_here/
|
iamnotdeadnuts
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsta8y
| false | null |
t3_1jsta8y
|
/r/LocalLLaMA/comments/1jsta8y/llama_4_is_here/
| false | false |
self
| 0 | null |
Improving Qwen2.5-7B QA Fine-tune: Precision, Recall, F1-Score
| 1 |
[removed]
| 2025-04-06T12:50:45 |
https://www.reddit.com/r/LocalLLaMA/comments/1jstebv/improving_qwen257b_qa_finetune_precision_recall/
|
blitzz000
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jstebv
| false | null |
t3_1jstebv
|
/r/LocalLLaMA/comments/1jstebv/improving_qwen257b_qa_finetune_precision_recall/
| false | false |
self
| 1 | null |
Anyone else in the Deepseek R2 Llama 4 scout distilled waiting room
| 15 |
With llama 4 scout being a small MoE how likely is it that Deepseek will create a distilled R2 on the platform.
| 2025-04-06T12:59:31 |
https://www.reddit.com/r/LocalLLaMA/comments/1jstk2y/anyone_else_in_the_deepseek_r2_llama_4_scout/
|
muhts
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jstk2y
| false | null |
t3_1jstk2y
|
/r/LocalLLaMA/comments/1jstk2y/anyone_else_in_the_deepseek_r2_llama_4_scout/
| false | false |
self
| 15 | null |
Small Llama4 on the way?
| 44 | 2025-04-06T13:02:27 |
https://www.reddit.com/r/LocalLLaMA/comments/1jstm9f/small_llama4_on_the_way/
|
ApprehensiveAd3629
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jstm9f
| false | null |
t3_1jstm9f
|
/r/LocalLLaMA/comments/1jstm9f/small_llama4_on_the_way/
| false | false | 44 |
{'enabled': False, 'images': [{'id': 'uSKaK7ueXiHG33KHHPMjv5lt-3tCQlQu3O7q_Xm8vPY', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/QEHnDECxQD3q-O2oAl66rIYqCq8OgSb_SIwGWVs--NA.jpg?width=108&crop=smart&auto=webp&s=edb78a33d9e415634accf7bedfc5e1a39fa3f818', 'width': 108}], 'source': {'height': 200, 'url': 'https://external-preview.redd.it/QEHnDECxQD3q-O2oAl66rIYqCq8OgSb_SIwGWVs--NA.jpg?auto=webp&s=8c4cc8332d8db01161fc74543a8337460ce457b4', 'width': 200}, 'variants': {}}]}
|
||
Llama 4 AI Models with a 10M Context Window Just Released
| 1 |
[removed]
| 2025-04-06T13:14:18 |
https://youtube.com/shorts/4LqlbybCGVM?si=Aa-kBvO3QFKv8f3y
|
Kind-Industry-609
|
youtube.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jstuds
| false | null |
t3_1jstuds
|
/r/LocalLLaMA/comments/1jstuds/llama_4_ai_models_with_a_10m_context_window_just/
| false | false |
default
| 1 | null |
Did LocalLLaMA just become an oxymoron?
| 1 |
[removed]
| 2025-04-06T13:19:46 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsty5s/did_localllama_just_become_an_oxymoron/
|
VegaKH
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsty5s
| false | null |
t3_1jsty5s
|
/r/LocalLLaMA/comments/1jsty5s/did_localllama_just_become_an_oxymoron/
| false | false |
self
| 1 | null |
Llama 4 with a 10M Context Window Just Released
| 1 |
[removed]
| 2025-04-06T13:20:14 |
https://youtube.com/shorts/4LqlbybCGVM?si=Aa-kBvO3QFKv8f3y
|
Kind-Industry-609
|
youtube.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1jstygr
| false | null |
t3_1jstygr
|
/r/LocalLLaMA/comments/1jstygr/llama_4_with_a_10m_context_window_just_released/
| false | false |
default
| 1 | null |
Best multilingual and coding model for MBP
| 1 |
[removed]
| 2025-04-06T13:52:25 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsuljk/best_multilingual_and_coding_model_for_mbp/
|
DeliciousApricot4686
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsuljk
| false | null |
t3_1jsuljk
|
/r/LocalLLaMA/comments/1jsuljk/best_multilingual_and_coding_model_for_mbp/
| false | false |
self
| 1 | null |
UPDATE: DeepSeek-R1 671B Works with LangChainβs MCP Adapters & LangGraphβs Bigtool!
| 1 |
[removed]
| 2025-04-06T14:24:21 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsv9vr/update_deepseekr1_671b_works_with_langchains_mcp/
|
lc19-
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsv9vr
| false | null |
t3_1jsv9vr
|
/r/LocalLLaMA/comments/1jsv9vr/update_deepseekr1_671b_works_with_langchains_mcp/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'QqSY3F9i2BgB-OdT_JpQr1vBqr2oq4spYNzkghHXwCM', 'resolutions': [], 'source': {'height': 64, 'url': 'https://external-preview.redd.it/4PgIzt2dsWk0hsH_pv6fTscUBf4LNxa8vUF1zyE23u0.jpg?auto=webp&s=adf334dabc58b5ccda405f20fe4d11f983c41fe9', 'width': 64}, 'variants': {}}]}
|
Peplrexity AI Pro 12 months
| 1 |
[removed]
| 2025-04-06T14:32:17 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsvg1c/peplrexity_ai_pro_12_months/
|
Sandy_Zimmerman
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsvg1c
| false | null |
t3_1jsvg1c
|
/r/LocalLLaMA/comments/1jsvg1c/peplrexity_ai_pro_12_months/
| false | false |
self
| 1 | null |
UPDATE: DeepSeek-R1 671B Works with LangChainβs MCP Adapters & LangGraphβs Bigtool!
| 5 |
I've just updated my GitHub repo with TWO new Jupyter Notebook tutorials showing DeepSeek-R1 671B (DeepSeek AI) working seamlessly with both LangChain's MCP Adapters library and LangGraph's Bigtool library! π
π πππ§π ππ‘ππ’π§'π¬ πππ ππππ©πππ«π¬ + ππππ©ππππ€-ππ ππππ
This notebook tutorial demonstrates that even without having DeepSeek-R1 671B fine-tuned for tool calling or even without using my Tool-Ahead-of-Time package (since LangChain's MCP Adapters library works by first converting tools in MCP servers into LangChain tools), MCP still works with DeepSeek-R1 671B (with DeepSeek-R1 671B as the client)! This is likely because DeepSeek-R1 671B is a reasoning model and also how the prompts are written in LangChain's MCP Adapters library.
π§° πππ§π ππ«ππ©π‘'π¬ ππ’π ππ¨π¨π₯ + ππππ©ππππ€-ππ ππππ
LangGraph's Bigtool library is a recently released library by LangGraph which helps AI to do tool calling from a large number of tools.
This notebook tutorial demonstrates that even without having DeepSeek-R1 671B fine-tuned for tool calling or even without using my Tool-Ahead-of-Time package, LangGraph's Bigtool library still works with DeepSeek-R1 671B. Again, this is likely because DeepSeek-R1 671B is a reasoning model and also how the prompts are written in LangGraph's Bigtool library.
π€ Why is this important? Because it shows how versatile DeepSeek-R1 671B truly is!
Check out my latest tutorials and please give my GitHub repo a star if this was helpful β
Python package:
https://github.com/leockl/tool-ahead-of-time
JavaScript (JavaScript Developer)/TypeScript package:
https://github.com/leockl/tool-ahead-of-time-ts (note: implementation support for using LangGraph's Bigtool library with DeepSeek-R1 671B was not included for the JavaScript/TypeScript package as there is currently no JavaScript/TypeScript support for the LangGraph's Bigtool library)
BONUS: From various socials, it appears the newly released Meta's Llama 4 models (Scout & Maverick) have disappointed a lot of people. Having said that, Scout & Maverick has tool calling support provided by the Llama team via LangChain's ChatOpenAI class.
| 2025-04-06T14:33:44 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsvh7e/update_deepseekr1_671b_works_with_langchains_mcp/
|
lc19-
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsvh7e
| false | null |
t3_1jsvh7e
|
/r/LocalLLaMA/comments/1jsvh7e/update_deepseekr1_671b_works_with_langchains_mcp/
| false | false |
self
| 5 |
{'enabled': False, 'images': [{'id': 'UsZmMZWIs8F6y4wuZiK2yL4sqd1GXz_hCtk2yWV7YdM', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/a8j5ZGZzDYBpcAlKjOjUwXcH1gESC29FJxzxZGpWXFI.jpg?width=108&crop=smart&auto=webp&s=b20fac72571f7dd416c6e83212e117f89606b01e', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/a8j5ZGZzDYBpcAlKjOjUwXcH1gESC29FJxzxZGpWXFI.jpg?width=216&crop=smart&auto=webp&s=37b8ee7b6ba61ef126c9e6482ed12f105a0fa851', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/a8j5ZGZzDYBpcAlKjOjUwXcH1gESC29FJxzxZGpWXFI.jpg?width=320&crop=smart&auto=webp&s=1f2fcf047e74a50828ca5485dcaa80cf84df44cd', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/a8j5ZGZzDYBpcAlKjOjUwXcH1gESC29FJxzxZGpWXFI.jpg?width=640&crop=smart&auto=webp&s=da1ff2649ab863b458b83fd583b19fde61adb28d', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/a8j5ZGZzDYBpcAlKjOjUwXcH1gESC29FJxzxZGpWXFI.jpg?width=960&crop=smart&auto=webp&s=1592414b45f794398a84ad418a1948656d7d6a83', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/a8j5ZGZzDYBpcAlKjOjUwXcH1gESC29FJxzxZGpWXFI.jpg?width=1080&crop=smart&auto=webp&s=a19fccf881c9702f7b9886904cfee3b600e3a85b', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/a8j5ZGZzDYBpcAlKjOjUwXcH1gESC29FJxzxZGpWXFI.jpg?auto=webp&s=1e87ffef00bfe5b0325086b42c8181b51667a5c4', 'width': 1200}, 'variants': {}}]}
|
Llama 4 Maverick surpassing Claude 3.7 Sonnet, under DeepSeek V3.1 according to Artificial Analysis
| 229 | 2025-04-06T14:59:47 |
TKGaming_11
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsw1x6
| false | null |
t3_1jsw1x6
|
/r/LocalLLaMA/comments/1jsw1x6/llama_4_maverick_surpassing_claude_37_sonnet/
| false | false | 229 |
{'enabled': True, 'images': [{'id': 'syJUOhgUuQcQve4UYbI1FJDNfHyiing3IFXMcebYQlc', 'resolutions': [{'height': 42, 'url': 'https://preview.redd.it/bybxcks0b8te1.png?width=108&crop=smart&auto=webp&s=1abc1800b7ff14013fd819d5a926b385260f2f14', 'width': 108}, {'height': 85, 'url': 'https://preview.redd.it/bybxcks0b8te1.png?width=216&crop=smart&auto=webp&s=513e1b1d36df91ec59846d981464e59a374ce954', 'width': 216}, {'height': 125, 'url': 'https://preview.redd.it/bybxcks0b8te1.png?width=320&crop=smart&auto=webp&s=baf31e2813832ed1ea663a71b24d2ef0e65fb781', 'width': 320}, {'height': 251, 'url': 'https://preview.redd.it/bybxcks0b8te1.png?width=640&crop=smart&auto=webp&s=956027562bb411e338688cbf361aa64681ffdaa5', 'width': 640}, {'height': 377, 'url': 'https://preview.redd.it/bybxcks0b8te1.png?width=960&crop=smart&auto=webp&s=dd258731bec29c7b3e1eab5373d6e23e15b3ba5e', 'width': 960}, {'height': 425, 'url': 'https://preview.redd.it/bybxcks0b8te1.png?width=1080&crop=smart&auto=webp&s=fffabfb9e4f55c3eac230e79457b1620582930de', 'width': 1080}], 'source': {'height': 1612, 'url': 'https://preview.redd.it/bybxcks0b8te1.png?auto=webp&s=84fc40d547f68e7433a2d130b9b05e7f3ed9ac92', 'width': 4096}, 'variants': {}}]}
|
|||
What is the best local LLM I can run with a RTX 5070 Ti?
| 0 |
Which local LLM would you recommend running and in what configuration? I also have 32GB of state memory.
I have been using this setup mostly for gaming and image generation so far, but also want to experiment with Local LLMs and audio generation models now as well
| 2025-04-06T15:04:45 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsw65o/what_is_the_best_local_llm_i_can_run_with_a_rtx/
|
qwed113
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsw65o
| false | null |
t3_1jsw65o
|
/r/LocalLLaMA/comments/1jsw65o/what_is_the_best_local_llm_i_can_run_with_a_rtx/
| false | false |
self
| 0 | null |
Nvidia Spark vs Framework PC vs 3x 3090 RTX Build
| 1 |
[removed]
| 2025-04-06T15:07:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsw812/nvidia_spark_vs_framework_pc_vs_3x_3090_rtx_build/
|
eskaroll
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsw812
| false | null |
t3_1jsw812
|
/r/LocalLLaMA/comments/1jsw812/nvidia_spark_vs_framework_pc_vs_3x_3090_rtx_build/
| false | false |
self
| 1 | null |
Does Llama.cpp support Unsloth's Dynamic 4bit quants?
| 5 |
Everytime I try to use the convert\_hf\_to\_gguf script to create GGUF from one of Unsloth's Dynamic 4bit Quants models, I get an error. I have not found any documentation stating Llama.cpp supports these models or doesn't support these models. Do I need to try a different approach?
(running win 11, llama.cpp built from latest source with Vulkan support, python 3.10)
`(python) PS C:\Users\gera\llms\QwQ-32B-unsloth-bnb-4bit> python ..\..\localLlama\llama.cpp\convert_hf_to_gguf.py .`
`INFO:hf-to-gguf:Loading model:`
`INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only`
`INFO:hf-to-gguf:Exporting model...`
`INFO:hf-to-gguf:gguf: loading model weight map from 'model.safetensors.index.json'`
`INFO:hf-to-gguf:gguf: loading model part 'model-00001-of-00005.safetensors'`
`Traceback (most recent call last):`
`File "C:\Users\gera\localLlama\llama.cpp\convert_hf_to_gguf.py", line 5511, in <module>`
`main()`
`File "C:\Users\gera\localLlama\llama.cpp\convert_hf_to_gguf.py", line 5505, in main`
`model_instance.write()`
`File "C:\Users\gera\localLlama\llama.cpp\convert_hf_to_gguf.py", line 440, in write`
`self.prepare_tensors()`
`File "C:\Users\gera\localLlama\llama.cpp\convert_hf_to_gguf.py", line 281, in prepare_tensors`
`for name, data_torch in chain(self.generate_extra_tensors(), self.get_tensors()):`
`File "C:\Users\gera\localLlama\llama.cpp\convert_hf_to_gguf.py", line 157, in get_tensors`
`ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu"))`
`safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLarge`
| 2025-04-06T15:07:54 |
https://www.reddit.com/r/LocalLLaMA/comments/1jsw8s7/does_llamacpp_support_unsloths_dynamic_4bit_quants/
|
Deputius
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1jsw8s7
| false | null |
t3_1jsw8s7
|
/r/LocalLLaMA/comments/1jsw8s7/does_llamacpp_support_unsloths_dynamic_4bit_quants/
| false | false |
self
| 5 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.