title
stringlengths 1
300
| score
int64 0
8.54k
| selftext
stringlengths 0
40k
| created
timestamp[ns]date 2023-04-01 04:30:41
2025-06-30 03:16:29
⌀ | url
stringlengths 0
878
| author
stringlengths 3
20
| domain
stringlengths 0
82
| edited
timestamp[ns]date 1970-01-01 00:00:00
2025-06-26 17:30:18
| gilded
int64 0
2
| gildings
stringclasses 7
values | id
stringlengths 7
7
| locked
bool 2
classes | media
stringlengths 646
1.8k
⌀ | name
stringlengths 10
10
| permalink
stringlengths 33
82
| spoiler
bool 2
classes | stickied
bool 2
classes | thumbnail
stringlengths 4
213
| ups
int64 0
8.54k
| preview
stringlengths 301
5.01k
⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Qwen3 training recap 🐦🔥
| 11 |
\[ Pre-training \]
\> 36T of text tokens (instead of 18T previously). For reference 1 epoch of Meta's dataset is 30T of text AND other modalities.
\> 3 stages pre-training:
1) 30T with 4k
2) 5T of science/math/code and reasoning data, no info on ctx length so maybe short CoT?
3) 1T of context extension to 32k (no ruler/helmet benchmark..)
\> 8 KV heads instead of 2 or 4 in Qwen 2 <7B.
\> No attention bias, and QK Norm (per head)
\> Nice MoEs (with global batch load balancing ofc)
\[ Post-training \]
\> Frontier model using RL with cold start and this « thinking mode fusion »
\> Smol model are using (data, not logit) distillation.
I really like how they use there previous generation of model to extract pdf data and generate synthetic data for code and math!
Also seems like this part from the model card sent earlier in r/LocalLLaMa didn't make it in the blogpost.. even more excited for the blog post and see what are this "optimization techniques" and scaling laws!
https://preview.redd.it/godl7w3ibnxe1.png?width=1184&format=png&auto=webp&s=bdf88138b2abc67f54b189fd4e803e990a14f431
| 2025-04-28T21:44:33 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka7ifs/qwen3_training_recap/
|
eliebakk
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka7ifs
| false | null |
t3_1ka7ifs
|
/r/LocalLLaMA/comments/1ka7ifs/qwen3_training_recap/
| false | false | 11 | null |
|
Quants are getting confusing
| 34 |
How come IQ4_NL is just 907 MB?
And why is there huge difference between sizes like IQ1_S is 1.15 GB while IQ1_M is 16.2 GB, I would expect them to be of "similar" size.
What am I missing, or there's something wrong with unsloth Qwen3 quants?
| 2025-04-28T21:46:15 |
blaz3d7
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka7jqc
| false | null |
t3_1ka7jqc
|
/r/LocalLLaMA/comments/1ka7jqc/quants_are_getting_confusing/
| false | false | 34 |
{'enabled': True, 'images': [{'id': 'AJO_mnVL-SJmLiLJyLQBqu4rVn2AUmbUgAgmmnl3k8s', 'resolutions': [{'height': 98, 'url': 'https://preview.redd.it/u9yxc0k0cnxe1.jpeg?width=108&crop=smart&auto=webp&s=201bf4944d00004a3442fadf3c870493369d5c74', 'width': 108}, {'height': 197, 'url': 'https://preview.redd.it/u9yxc0k0cnxe1.jpeg?width=216&crop=smart&auto=webp&s=d46390fe82f94b2068555cf289c1fd1a0bdc668c', 'width': 216}, {'height': 292, 'url': 'https://preview.redd.it/u9yxc0k0cnxe1.jpeg?width=320&crop=smart&auto=webp&s=d3d00db56102a3bb8b91938db10f92bb42f1a05a', 'width': 320}, {'height': 585, 'url': 'https://preview.redd.it/u9yxc0k0cnxe1.jpeg?width=640&crop=smart&auto=webp&s=5005d41520c34013547cb2c39799384f43fc3c58', 'width': 640}, {'height': 878, 'url': 'https://preview.redd.it/u9yxc0k0cnxe1.jpeg?width=960&crop=smart&auto=webp&s=c6f2bf6831e654da7a1174df227dad2d95907f2e', 'width': 960}, {'height': 988, 'url': 'https://preview.redd.it/u9yxc0k0cnxe1.jpeg?width=1080&crop=smart&auto=webp&s=28e491f0313977985e06eee995deb8cb73e2ccdf', 'width': 1080}], 'source': {'height': 1117, 'url': 'https://preview.redd.it/u9yxc0k0cnxe1.jpeg?auto=webp&s=383d2180064b4b8a44de54ad05b3c80ed0234fe0', 'width': 1220}, 'variants': {}}]}
|
||
Qwen 3 Finetunes
| 4 |
With how much hype is around Qwen3, what kind of finetunes are you all expecting for this model?
I have a couple projects in mind... the think mode is gonna come in handy for those.
| 2025-04-28T21:48:32 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka7lh4/qwen_3_finetunes/
|
Reader3123
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka7lh4
| false | null |
t3_1ka7lh4
|
/r/LocalLLaMA/comments/1ka7lh4/qwen_3_finetunes/
| false | false |
self
| 4 | null |
Why doesn’t multi-GPU actually speed up LLM inference?
| 3 |
Hi everyone,
I keep reading “multi-GPU doesn’t really help inference latency,” and see it in benchmarks. But when I crunch the numbers I *still* expect a solid speed-up. Maybe I’m missing something obvious, so I'd love to hear what you think.
# My toy setup :
**Model**: 7B parameters (i.e. llama 7b), decoder-only, 32 layers, d = 4096, FP16
**GPUS**: two identical A100-40 GB (312 TFLOPS FP16, 1.555 TB/s HBM, connected by NVLink).
**Parallelism plan**: split the stack in half (16 layers on GPU-0, 16 on GPU-1) → classic 2-stage pipeline
Single-GPU numbers I trust :
Mem bandwidth for A100 = 1555 GB/s = 1.555 × 10¹² bytes/s
A100 peak compute (FP16 Tensor-Core) = 312 TFLOPS = 312 × 10¹² FLOP/s
N = 7 × 10⁹ parameters
P (weight size) = N × 2 bytes/param = 14 × 10⁹ bytes
pure compute cost per **one token**
2 × N (add + mul) / A100 peak compute
(2 × 7 × 10⁹) / (312 × 10¹²) = 4.49 × 10⁻⁵ s
To load all weights in mem we need
P / A100 mem bandwidth
(14 × 10⁹) / (1.555 × 10¹²) = 9.01 × 10⁻³ s ≈ 9.01 ms
We ignore KV‑cache traffic, MBU, Kernel/NVLink overhead and tiny activations.
If you are interested to deep dive, here is a good blog post : [https://kipp.ly/transformer-inference-arithmetic/](https://kipp.ly/transformer-inference-arithmetic/)
**Because of that we are memory bandwidth bound.**
=> TPOT (memory-bound) dominated by 9 ms
# Naïve expectation for two GPUs (A & B)
* Each stage now loads only **7 GB.**
* The best way to do that would be to overlap, so after the pipeline is full I *think* a new token should pop out every **\~4.5 ms** instead of 9 ms (2 × higher tok/s): When GPU B is loading weigths for generation of token 1, GPU A starts loading weights for generation of token 2.
But in every benchmark I see it's not the case. Is it from bad dynamic GPU orchestration ? I.e. we do not overlap \[when GPU 1 finishes it waits for GPU 2 to start loading weights (remember as we are memory bound)\] ? Are PyTorch / HF PP wrappers just bad at keeping both devices saturated?
I came to the conclusion that most off-the-shelf PP schedulers (PyTorch PP, HF Accelerate, DeepSpeed inference) run the decode stage with **exactly one micro-batch**. So no overlap happens. Why ?
Huge thanks for any pointers, corrections or additional discussion.
| 2025-04-28T21:56:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka7s3q/why_doesnt_multigpu_actually_speed_up_llm/
|
regis_lekeuf
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka7s3q
| false | null |
t3_1ka7s3q
|
/r/LocalLLaMA/comments/1ka7s3q/why_doesnt_multigpu_actually_speed_up_llm/
| false | false |
self
| 3 | null |
On chat.qwen site when using 235b model like it's non thinking and thinking model too right like when we click on thinking option it uses same 235b model to think
| 1 | 2025-04-28T21:58:09 |
Independent-Wind4462
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka7tm6
| false | null |
t3_1ka7tm6
|
/r/LocalLLaMA/comments/1ka7tm6/on_chatqwen_site_when_using_235b_model_like_its/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'd53Koys4srJS808pkNfXbSjp-ukY0nZj7BZnaK5sytU', 'resolutions': [{'height': 60, 'url': 'https://preview.redd.it/5yva6zo4enxe1.jpeg?width=108&crop=smart&auto=webp&s=df0aa6f18c69f3fb57959e85aa822ceb14e08223', 'width': 108}, {'height': 121, 'url': 'https://preview.redd.it/5yva6zo4enxe1.jpeg?width=216&crop=smart&auto=webp&s=0adb204ecfec7bdcf5270ca1502df5c813c0c155', 'width': 216}, {'height': 180, 'url': 'https://preview.redd.it/5yva6zo4enxe1.jpeg?width=320&crop=smart&auto=webp&s=8fd06c19917df65b67b52f9136dc5a2a076c96d8', 'width': 320}, {'height': 360, 'url': 'https://preview.redd.it/5yva6zo4enxe1.jpeg?width=640&crop=smart&auto=webp&s=d9aa971b3cb60fb621c76849eac151e391867b74', 'width': 640}, {'height': 540, 'url': 'https://preview.redd.it/5yva6zo4enxe1.jpeg?width=960&crop=smart&auto=webp&s=91d913a4fcf84754850a41742efa42cba3fbaec9', 'width': 960}, {'height': 607, 'url': 'https://preview.redd.it/5yva6zo4enxe1.jpeg?width=1080&crop=smart&auto=webp&s=34cc924c5d8cbb51c79765c6c36c28954ea10bf3', 'width': 1080}], 'source': {'height': 1920, 'url': 'https://preview.redd.it/5yva6zo4enxe1.jpeg?auto=webp&s=8aac95b5fc2eddb4601322d6bc2cc15dbce7b4f6', 'width': 3413}, 'variants': {}}]}
|
|||
Local LLM that supports openAI API tool call format
| 1 |
[removed]
| 2025-04-28T22:03:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka7y33/local_llm_that_supports_openai_api_tool_call/
|
Wooden_Yam1924
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka7y33
| false | null |
t3_1ka7y33
|
/r/LocalLLaMA/comments/1ka7y33/local_llm_that_supports_openai_api_tool_call/
| false | false |
self
| 1 | null |
A Developer’s Guide to Build Your Local OpenAI Operator on macOS
| 1 |
[removed]
| 2025-04-28T22:08:16 |
https://x.com/trycua/status/1916890472901169597
|
sandropuppo
|
x.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka823p
| false | null |
t3_1ka823p
|
/r/LocalLLaMA/comments/1ka823p/a_developers_guide_to_build_your_local_openai/
| false | false | 1 |
{'enabled': False, 'images': [{'id': 'rapiLjWwRrtONxICwWojvYZILFYVcNOf6-SJS-tosf8', 'resolutions': [{'height': 68, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=108&crop=smart&auto=webp&s=d9e098d0cc21772694bbc73dab98c93716a59085', 'width': 108}, {'height': 137, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=216&crop=smart&auto=webp&s=4b4859b38aa58353ab34895c6ba98719ab8088d2', 'width': 216}, {'height': 204, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=320&crop=smart&auto=webp&s=fdb466f43c3c90eab81aa630a53e3a116e369423', 'width': 320}, {'height': 408, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=640&crop=smart&auto=webp&s=4b69854ac82a9f742f757ba519b19b0814819fba', 'width': 640}, {'height': 612, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=960&crop=smart&auto=webp&s=59dc184fa665fb8e4772c0cf487b1073bb027ee0', 'width': 960}, {'height': 688, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=1080&crop=smart&auto=webp&s=2e362d95604db3a6fdcd1f085d05c659f1fb4369', 'width': 1080}], 'source': {'height': 825, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?auto=webp&s=ff092b442b37981929e1e714717a5cdd93f57b15', 'width': 1294}, 'variants': {}}]}
|
|
3090 Alternatives?
| 1 |
[removed]
| 2025-04-28T22:10:39 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka83x8/3090_alternatives/
|
Beneficial-Cup-2969
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka83x8
| false | null |
t3_1ka83x8
|
/r/LocalLLaMA/comments/1ka83x8/3090_alternatives/
| false | false |
self
| 1 | null |
Qwen3 speculative decoding tips, ideas, benchmarks, questions generic thread.
| 11 |
Qwen3 speculative decoding tips, ideas, benchmarks, questions generic thread.
To start some questions:
I see that Qwen3-4B, Qwen3-1.7B, Qwen3-0.6B are simply listed in the blog as having 32k context length vs. the larger models having 128k.
So to what extent does that impair their use as draft models if you're using the large model with long-ish context e.g. 32k or over?
Maybe the small context 'local' statistics tend to be overwhelming in most cases to predict the next token so perhaps it wouldn't deteriorate the predictive accuracy much to have a draft context length limit of much less than the full model? I'm guessing this has already been benchmarked and a "rule of thumb" about draft context sufficiency has come out?
Also I wonder how the Qwen3-30B-A3B model could potentially fare in the role of a draft model for Qwen3-32B, Qwen3-235B-A22B? Is it not a plausibly reasonable idea for some structural / model specific reason?
Anyway how's speculation working so far for those who have started benchmarking these for various use cases (text, coding in XYZ language, ...)?
| 2025-04-28T22:12:55 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka85pe/qwen3_speculative_decoding_tips_ideas_benchmarks/
|
Calcidiol
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka85pe
| false | null |
t3_1ka85pe
|
/r/LocalLLaMA/comments/1ka85pe/qwen3_speculative_decoding_tips_ideas_benchmarks/
| false | false |
self
| 11 | null |
Alternative GPUs for LLM Model Training (vs. 3090)?
| 1 |
[removed]
| 2025-04-28T22:16:59 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka88ys/alternative_gpus_for_llm_model_training_vs_3090/
|
Beneficial-Cup-2969
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka88ys
| false | null |
t3_1ka88ys
|
/r/LocalLLaMA/comments/1ka88ys/alternative_gpus_for_llm_model_training_vs_3090/
| false | false |
self
| 1 | null |
New to fine-tuning pytorch or tensorflow?
| 0 |
Hey folks,
Im new to fine tuning and wanted to start messing around with LLM fine-tuning.
Looks like PyTorch and TensorFlow are the main ways any advice or experiences to share to help me get started?
Appreciate it
| 2025-04-28T22:17:04 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka891o/new_to_finetuning_pytorch_or_tensorflow/
|
toolhouseai
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka891o
| false | null |
t3_1ka891o
|
/r/LocalLLaMA/comments/1ka891o/new_to_finetuning_pytorch_or_tensorflow/
| false | false |
self
| 0 | null |
Qwen3-30B-A3B is what most people have been waiting for
| 926 |
A QwQ competitor that limits its thinking that uses MoE with very small experts for lightspeed inference.
It's out, it's the real deal, Q5 is competing with QwQ easily in my personal local tests and pipelines.
No excuse now - intelligence that used to be SOTA now runs on modest gaming rigs - *GO BUILD SOMETHING COOL*
| 2025-04-28T22:19:39 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka8b2u/qwen330ba3b_is_what_most_people_have_been_waiting/
|
ForsookComparison
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8b2u
| false | null |
t3_1ka8b2u
|
/r/LocalLLaMA/comments/1ka8b2u/qwen330ba3b_is_what_most_people_have_been_waiting/
| false | false |
self
| 926 | null |
Qwen 3: unimpressive coding performance so far
| 95 |
Jumping ahead of the classic "OMG QWEN 3 IS THE LITERAL BEST IN EVERYTHING" and providing a small feedback on it's coding characteristics.
TECHNOLOGIES USED:
.NET 9
Typescript
React 18
Material UI.
MODEL USED:
Qwen3-235B-A22B (From Qwen AI chat)
PROMPTS (Void of code because it's a private project):
\- "My current code shows for a split second that \[RELEVANT\_DATA\] is missing, only to then display \[RELEVANT\_DATA\]properly. I do not want that split second missing warning to happen."
RESULT: Fairly insignificant code change suggestions that did not fix the problem, when prompted that the solution was not successful and the rendering issue persisted, it repeated the same code again.
\- "Please split $FAIRLY\_BIG\_DOTNET\_CLASS (Around 3K lines of code) into smaller classes to enhance readability and maintainability"
RESULT: Code was *mostly* correct, but it really hallucinated some stuff and threw away some other without a specific reason.
So yeah, this is a very hot opinion about Qwen 3
**THE PROS**
Follows instruction, doesn't spit out ungodly amount of code like Gemini Pro 2.5 does, fairly fast (at least on chat I guess)
**THE CONS**
Not so amazing coding performance, I'm sure a coder variant will fare much better though
| 2025-04-28T22:19:57 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka8ban/qwen_3_unimpressive_coding_performance_so_far/
|
ps5cfw
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8ban
| false | null |
t3_1ka8ban
|
/r/LocalLLaMA/comments/1ka8ban/qwen_3_unimpressive_coding_performance_so_far/
| false | false |
self
| 95 | null |
We are running Qwen3 30b (2 H100 replicas) and Qwen 235b and (4xh200 Replicas)
| 1 |
[removed]
| 2025-04-28T22:21:54 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka8cw5/we_are_running_qwen3_30b_2_h100_replicas_and_qwen/
|
parasail_io
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8cw5
| false | null |
t3_1ka8cw5
|
/r/LocalLLaMA/comments/1ka8cw5/we_are_running_qwen3_30b_2_h100_replicas_and_qwen/
| false | false |
self
| 1 | null |
Qwen3 32B FP8 memory + vllm?
| 1 |
Am I crazy/my math is wrong or should Qwen3-32B-FP8 fit in ~21GB of vram? I'm currently getting CUDA OOM with vLLM (2x3060):
```
docker run \
--name my_vllm_container \
--gpus '"device=0,1"' \
-v /mnt/models:/root/models \
-p 8000:8000 \
--ipc=host \
vllm/vllm-openai:latest \
--model /root/models/Qwen3-32B-FP8 \
--served-model-name Qwen/Qwen3-32B-FP8 \
--gpu-memory-utilization 1 \
--pipeline-parallel-size 2 \
--max-num-seqs 2 \
--max-model-len 2292 \
--block-size 32 \
--max-num-batched-tokens 2292 \
--enable-reasoning \
--reasoning-parser deepseek_r1
```
(Yes I'm aware that the model itself won't quite run yet, waiting on the new vllm docker image to go live in a few hours. Mostly just trying to get past this CUDA OOM, which I can on my 2x4090)
| 2025-04-28T22:27:16 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka8h6u/qwen3_32b_fp8_memory_vllm/
|
tddammo1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8h6u
| false | null |
t3_1ka8h6u
|
/r/LocalLLaMA/comments/1ka8h6u/qwen3_32b_fp8_memory_vllm/
| false | false |
self
| 1 | null |
Qwen3-30B-A3B is magic.
| 241 |
I don't believe a model this good runs at 20 tps on my 4gb gpu (rx 6550m).
Running it through paces, seems like the benches were right on.
| 2025-04-28T22:34:23 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka8n18/qwen330ba3b_is_magic/
|
thebadslime
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8n18
| false | null |
t3_1ka8n18
|
/r/LocalLLaMA/comments/1ka8n18/qwen330ba3b_is_magic/
| false | false |
self
| 241 | null |
Why is a <9 GB file on my pc able to do this? Qwen 3 14B Q4_K_S one shot prompt: "give me a snake html game, fully working"
| 173 | 2025-04-28T22:36:27 |
https://v.redd.it/fvl50yhjknxe1
|
sebastianmicu24
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8omw
| false |
{'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/fvl50yhjknxe1/DASHPlaylist.mpd?a=1748471802%2CMDVhY2ZmYzM3Mjc5YWIzN2NiOTUwYzQyOWU0MGZjYTc4NzY2YzFjYmQ2NDZhYzk5ZmEwMWZkYjIyMTdhOGM4OQ%3D%3D&v=1&f=sd', 'duration': 25, 'fallback_url': 'https://v.redd.it/fvl50yhjknxe1/DASH_720.mp4?source=fallback', 'has_audio': True, 'height': 714, 'hls_url': 'https://v.redd.it/fvl50yhjknxe1/HLSPlaylist.m3u8?a=1748471802%2CNDU0YmMwM2VmMGVhZWE5YTY4OGYwYjNlYWIxNWJkOTlkMDQ4ZGY2MjdkYjE5NWI2YjRiNzU2YTgwNmU4OWJhMQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/fvl50yhjknxe1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1280}}
|
t3_1ka8omw
|
/r/LocalLLaMA/comments/1ka8omw/why_is_a_9_gb_file_on_my_pc_able_to_do_this_qwen/
| false | false | 173 |
{'enabled': False, 'images': [{'id': 'NDJxdDh6aGprbnhlMYpX_8uwwmYorZvO9YWXhNvaB0JdvRfyxdOqaIayhC1C', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/NDJxdDh6aGprbnhlMYpX_8uwwmYorZvO9YWXhNvaB0JdvRfyxdOqaIayhC1C.png?width=108&crop=smart&format=pjpg&auto=webp&s=a403efc7805a2704089cc77ee143a4fa2ab06226', 'width': 108}, {'height': 120, 'url': 'https://external-preview.redd.it/NDJxdDh6aGprbnhlMYpX_8uwwmYorZvO9YWXhNvaB0JdvRfyxdOqaIayhC1C.png?width=216&crop=smart&format=pjpg&auto=webp&s=70f57005f36bf0c2491c52e5598ecf8c2c09d492', 'width': 216}, {'height': 178, 'url': 'https://external-preview.redd.it/NDJxdDh6aGprbnhlMYpX_8uwwmYorZvO9YWXhNvaB0JdvRfyxdOqaIayhC1C.png?width=320&crop=smart&format=pjpg&auto=webp&s=f09dac479365058b2f105bb7f61605035340ac7e', 'width': 320}, {'height': 357, 'url': 'https://external-preview.redd.it/NDJxdDh6aGprbnhlMYpX_8uwwmYorZvO9YWXhNvaB0JdvRfyxdOqaIayhC1C.png?width=640&crop=smart&format=pjpg&auto=webp&s=f3d2b73509fd7737a1b58a6d69988a7e80666932', 'width': 640}, {'height': 536, 'url': 'https://external-preview.redd.it/NDJxdDh6aGprbnhlMYpX_8uwwmYorZvO9YWXhNvaB0JdvRfyxdOqaIayhC1C.png?width=960&crop=smart&format=pjpg&auto=webp&s=73db7186b0416e6b7ef4f96f38f614e45263f1ae', 'width': 960}, {'height': 603, 'url': 'https://external-preview.redd.it/NDJxdDh6aGprbnhlMYpX_8uwwmYorZvO9YWXhNvaB0JdvRfyxdOqaIayhC1C.png?width=1080&crop=smart&format=pjpg&auto=webp&s=4d175dfa007c84db52e07c856eb157c1b3d01ff4', 'width': 1080}], 'source': {'height': 1072, 'url': 'https://external-preview.redd.it/NDJxdDh6aGprbnhlMYpX_8uwwmYorZvO9YWXhNvaB0JdvRfyxdOqaIayhC1C.png?format=pjpg&auto=webp&s=ced794efeaa583db82cbd3151444edc802f137d6', 'width': 1920}, 'variants': {}}]}
|
||
Qwen didn't just cook. They had a whole barbecue!
| 1,203 | 2025-04-28T22:39:32 |
MushroomGecko
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8r2t
| false | null |
t3_1ka8r2t
|
/r/LocalLLaMA/comments/1ka8r2t/qwen_didnt_just_cook_they_had_a_whole_barbecue/
| false | false |
default
| 1,203 |
{'enabled': True, 'images': [{'id': 'z8nu50rdlnxe1', 'resolutions': [{'height': 106, 'url': 'https://preview.redd.it/z8nu50rdlnxe1.jpeg?width=108&crop=smart&auto=webp&s=a0892ca1a7e0c789eb193b6445168eca124cc7c8', 'width': 108}, {'height': 212, 'url': 'https://preview.redd.it/z8nu50rdlnxe1.jpeg?width=216&crop=smart&auto=webp&s=1cc0ac71cf3616fa464a503ed3ad6b656de4bb97', 'width': 216}, {'height': 314, 'url': 'https://preview.redd.it/z8nu50rdlnxe1.jpeg?width=320&crop=smart&auto=webp&s=9393acdfc60bee35b04d185957c2999651e0c875', 'width': 320}, {'height': 628, 'url': 'https://preview.redd.it/z8nu50rdlnxe1.jpeg?width=640&crop=smart&auto=webp&s=42135693763896c9c446089c6fa47afc509ad7f4', 'width': 640}, {'height': 943, 'url': 'https://preview.redd.it/z8nu50rdlnxe1.jpeg?width=960&crop=smart&auto=webp&s=29e506300edc394cda43309e61887516a59221e3', 'width': 960}, {'height': 1060, 'url': 'https://preview.redd.it/z8nu50rdlnxe1.jpeg?width=1080&crop=smart&auto=webp&s=6d39f1bb3d81c384e514add5bc1c535330ab23c9', 'width': 1080}], 'source': {'height': 1167, 'url': 'https://preview.redd.it/z8nu50rdlnxe1.jpeg?auto=webp&s=4fcfc89b120c721d57762f9c634b6e6cd7d78567', 'width': 1188}, 'variants': {}}]}
|
||
Ollama /api/chat to /v1/chat/completions proxy
| 1 |
Hi all, does anyone have or know of a lightweight proxy that would accept requests for Ollama's /api/chat endpoint and proxy them to an openai compatible /v1/chat/completions endpoint, returning an Ollama ChatResponse to the calling client?
This may seem like an weird request, but there is an app not under my control that I use that makes all of its requests to Ollama's /api/chat and I want to use vLLM or something other than Ollama without making changes to the app.
| 2025-04-28T22:41:35 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka8sn9/ollama_apichat_to_v1chatcompletions_proxy/
|
databasehead
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8sn9
| false | null |
t3_1ka8sn9
|
/r/LocalLLaMA/comments/1ka8sn9/ollama_apichat_to_v1chatcompletions_proxy/
| false | false |
self
| 1 | null |
A Developer’s Guide to Build Your Local OpenAI Operator on macOS
| 1 |
If you’re poking around with OpenAI Operator on Apple Silicon (or just want to build AI agents that can actually use a computer like a human), this is for you. I've written a guide to walk you through getting started with cua-agent, show you how to pick the right model/loop for your use case, and share some code patterns that’ll get you up and running fast.
Here is the full guide, but you can find a sneakpeek below: [https://x.com/trycua/status/1916890472901169597](https://x.com/trycua/status/1916890472901169597)
# What is cua-agent?
Think of `cua-agent` as the toolkit that lets you skip the gnarly boilerplate of screenshotting, sending context to an LLM, parsing its output, and safely running actions in a VM. It gives you a clean Python API for building “Computer-Use Agents” (CUAs) that can click, type, and see what’s on the screen. You can swap between OpenAI, Anthropic, UI-TARS, or local open-source models (Ollama, LM Studio, vLLM, etc.) with almost zero code changes.
# Setup: Get Rolling in 5 Minutes
**Prereqs:**
* Python 3.10+ (Conda or venv is fine)
* macOS CUA image already set up (see Part 1 if you haven’t)
* API keys for OpenAI/Anthropic (optional if you want to use local models)
* Ollama installed if you want to run local models
**Install everything:**
bashpip install "cua-agent[all]"
Or cherry-pick what you need:
bashpip install "cua-agent[openai]"
# OpenAI
pip install "cua-agent[anthropic]"
# Anthropic
pip install "cua-agent[uitars]"
# UI-TARS
pip install "cua-agent[omni]"
# Local VLMs
pip install "cua-agent[ui]"
# Gradio UI
**Set up your Python environment:**
bashconda create -n cua-agent python=3.10
conda activate cua-agent
# or
python -m venv cua-env
source cua-env/bin/activate
**Export your API keys:**
bashexport OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...
# Agent Loops: Which Should You Use?
Here’s the quick-and-dirty rundown:
|Loop|Models it Runs|When to Use It|
|:-|:-|:-|
||
|`OPENAI`|OpenAI CUA Preview|Browser tasks, best web automation, Tier 3 only|
|`ANTHROPIC`|Claude 3.5/3.7|Reasoning-heavy, multi-step, robust workflows|
|`UITARS`|UI-TARS-1.5 (ByteDance)|OS/desktop automation, low latency, local|
|`OMNI`|Any VLM (Ollama, etc.)|Local, open-source, privacy/cost-sensitive|
**TL;DR:**
* Use `OPENAI` for browser stuff if you have access.
* Use `UITARS` for desktop/OS automation.
* Use `OMNI` if you want to run everything locally or avoid API costs.
# Your First Agent in ~15 Lines
pythonimport asyncio
from computer import Computer
from agent import ComputerAgent, LLMProvider, LLM, AgentLoop
async def main():
async with Computer() as macos:
agent = ComputerAgent(
computer=macos,
loop=AgentLoop.OPENAI,
model=LLM(provider=LLMProvider.OPENAI)
)
task = "Open Safari and search for 'Python tutorials'"
async for result in agent.run(task):
print(result.get('text'))
if __name__ == "__main__":
asyncio.run(main())
Just drop that in a file and run it. The agent will spin up a VM, open Safari, and run your task. No need to handle screenshots, parsing, or retries yourself1.
# Chaining Tasks: Multi-Step Workflows
You can feed the agent a list of tasks, and it’ll keep context between them:
pythontasks = [
"Open Safari and go to github.com",
"Search for 'trycua/cua'",
"Open the repository page",
"Click on the 'Issues' tab",
"Read the first open issue"
]
for i, task in enumerate(tasks):
print(f"\nTask {i+1}/{len(tasks)}: {task}")
async for result in agent.run(task):
print(f" → {result.get('text')}")
print(f"✅ Task {i+1} done")
Great for automating actual workflows, not just single clicks1.
# Local Models: Save Money, Run Everything On-Device
Want to avoid OpenAI/Anthropic API costs? You can run agents with open-source models locally using Ollama, LM Studio, vLLM, etc.
**Example:**
bashollama pull gemma3:4b-it-q4_K_M
pythonagent = ComputerAgent(
computer=macos_computer,
loop=AgentLoop.OMNI,
model=LLM(
provider=LLMProvider.OLLAMA,
name="gemma3:4b-it-q4_K_M"
)
)
You can also point to any OpenAI-compatible endpoint (LM Studio, vLLM, LocalAI, etc.)1.
# Debugging & Structured Responses
Every action from the agent gives you a rich, structured response:
* Action text
* Token usage
* Reasoning trace
* Computer action details (type, coordinates, text, etc.)
This makes debugging and logging a breeze. Just print the result dict or log it to a file for later inspection1.
# Visual UI (Optional): Gradio
If you want a UI for demos or quick testing:
pythonfrom agent.ui.gradio.app import create_gradio_ui
if __name__ == "__main__":
app = create_gradio_ui()
app.launch(share=False)
# Local only
Supports model/loop selection, task input, live screenshots, and action history.
Set `share=True` for a public link (with optional password)1.
# Tips & Gotchas
* You can swap loops/models with almost no code changes.
* Local models are great for dev, testing, or privacy.
* `.gradio_settings.json` saves your UI config-add it to `.gitignore`.
* For UI-TARS, deploy locally or on Hugging Face and use OAICOMPAT provider.
* Check the structured response for debugging, not just the action text.
| 2025-04-28T22:43:42 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka8ubt/a_developers_guide_to_build_your_local_openai/
|
sandropuppo
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8ubt
| false | null |
t3_1ka8ubt
|
/r/LocalLLaMA/comments/1ka8ubt/a_developers_guide_to_build_your_local_openai/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'rapiLjWwRrtONxICwWojvYZILFYVcNOf6-SJS-tosf8', 'resolutions': [{'height': 68, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=108&crop=smart&auto=webp&s=d9e098d0cc21772694bbc73dab98c93716a59085', 'width': 108}, {'height': 137, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=216&crop=smart&auto=webp&s=4b4859b38aa58353ab34895c6ba98719ab8088d2', 'width': 216}, {'height': 204, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=320&crop=smart&auto=webp&s=fdb466f43c3c90eab81aa630a53e3a116e369423', 'width': 320}, {'height': 408, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=640&crop=smart&auto=webp&s=4b69854ac82a9f742f757ba519b19b0814819fba', 'width': 640}, {'height': 612, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=960&crop=smart&auto=webp&s=59dc184fa665fb8e4772c0cf487b1073bb027ee0', 'width': 960}, {'height': 688, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?width=1080&crop=smart&auto=webp&s=2e362d95604db3a6fdcd1f085d05c659f1fb4369', 'width': 1080}], 'source': {'height': 825, 'url': 'https://external-preview.redd.it/FUhBoJIgme7bEvj5fUw4wpAwnrmHhgRoG1tLJhJcMcg.jpg?auto=webp&s=ff092b442b37981929e1e714717a5cdd93f57b15', 'width': 1294}, 'variants': {}}]}
|
A Developer’s Guide to Build Your Local OpenAI Operator on macOS
| 1 |
[removed]
| 2025-04-28T22:44:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka8uu0/a_developers_guide_to_build_your_local_openai/
|
sandropuppo
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka8uu0
| false | null |
t3_1ka8uu0
|
/r/LocalLLaMA/comments/1ka8uu0/a_developers_guide_to_build_your_local_openai/
| false | false |
self
| 1 | null |
Qwen 3 + KTransformers 0.3 (+AMX) = AI Workstation/PC
| 37 |
Qwen 3 is out, and so is KTransformers v0.3!
Thanks to the great support from the Qwen team, we're excited to announce that KTransformers now supports Qwen3MoE from day one.
We're also taking this opportunity to open-source long-awaited AMX support in KTransformers!
One thing that really excites me about Qwen3MoE is how it \*\***targets the sweet spots**\*\* for both local workstations and consumer PCs, compared to massive models like the 671B giant.
Specifically, Qwen3MoE offers two different sizes: 235B-A22 and 30B-A3B, both designed to better fit real-world setups.
We ran tests in two typical scenarios:
\- (1) Server-grade CPU (Xeon4) + 4090
\- (2) Consumer-grade CPU (Core i9-14900KF + dual-channel 4000MT) + 4090
The results are very promising!
https://preview.redd.it/hr7iabtfonxe1.png?width=2879&format=png&auto=webp&s=1f2c40938b4bb6cf8799fd6ca86d4bec89092c3e
https://preview.redd.it/roilfwgionxe1.png?width=783&format=png&auto=webp&s=0f28d11d8d7b6d4ba4473574fd6816811022e8f5
Enjoy the new release — and stay tuned for even more exciting updates coming soon!
To help understand our AMX optimization, we also provide a following document: [https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/AMX.md](https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/AMX.md)
| 2025-04-28T22:57:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka94qx/qwen_3_ktransformers_03_amx_ai_workstationpc/
|
CombinationNo780
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka94qx
| false | null |
t3_1ka94qx
|
/r/LocalLLaMA/comments/1ka94qx/qwen_3_ktransformers_03_amx_ai_workstationpc/
| false | false | 37 |
{'enabled': False, 'images': [{'id': 'qQG1dXI1kVnCq7MXfQFY49GCAbw9ktZA4xrMOwT6EWQ', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/GE12j_uxYOg7mq2etbhk8FlpkSr4OaQWJJb8HkAe49M.jpg?width=108&crop=smart&auto=webp&s=778301cec7c1aef3bd0dd7dd4449496a3c30172b', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/GE12j_uxYOg7mq2etbhk8FlpkSr4OaQWJJb8HkAe49M.jpg?width=216&crop=smart&auto=webp&s=d2b25fcef6f05a51bbacbec8ab3fe92548ba9ae3', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/GE12j_uxYOg7mq2etbhk8FlpkSr4OaQWJJb8HkAe49M.jpg?width=320&crop=smart&auto=webp&s=dc62ad023d9f77ae5545f8391f8d08faba83d6b8', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/GE12j_uxYOg7mq2etbhk8FlpkSr4OaQWJJb8HkAe49M.jpg?width=640&crop=smart&auto=webp&s=0027d7924c1f60a8126ae742bbf4b9f5f9abf8fe', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/GE12j_uxYOg7mq2etbhk8FlpkSr4OaQWJJb8HkAe49M.jpg?width=960&crop=smart&auto=webp&s=8d5d55865336c20bcb845704e0c29130d3340f07', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/GE12j_uxYOg7mq2etbhk8FlpkSr4OaQWJJb8HkAe49M.jpg?width=1080&crop=smart&auto=webp&s=96f0df9d13427849ae6f42ffe95a135149678ae2', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/GE12j_uxYOg7mq2etbhk8FlpkSr4OaQWJJb8HkAe49M.jpg?auto=webp&s=a1cdc165760e7a8eff630f74f9ffa9a0306d79e1', 'width': 1200}, 'variants': {}}]}
|
|
Qwen3-30B-A3B runs at 130 tokens-per-second prompt processing and 60 tokens-per-second generation speed on M1 Max
| 68 |
https://reddit.com/link/1ka9cp2/video/ra5xmwg5pnxe1/player
This thing freaking *rips*
| 2025-04-28T23:07:27 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka9cp2/qwen330ba3b_runs_at_130_tokenspersecond_prompt/
|
mark-lord
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka9cp2
| false | null |
t3_1ka9cp2
|
/r/LocalLLaMA/comments/1ka9cp2/qwen330ba3b_runs_at_130_tokenspersecond_prompt/
| false | false |
self
| 68 | null |
Qwen3-32B-GGUF Q_5_S fits neatly on 24 GB cards.
| 7 |
The tittle says it all. A few days ago a post about GLM-4-32B Q5\_K\_S working well on 24 GB cards was quite popular.
Qwen 3 works just as well. I'm getting about 10 tokens/s on a 3090 using Ollama.
| 2025-04-28T23:13:47 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka9hpl/qwen332bgguf_q_5_s_fits_neatly_on_24_gb_cards/
|
atape_1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka9hpl
| false | null |
t3_1ka9hpl
|
/r/LocalLLaMA/comments/1ka9hpl/qwen332bgguf_q_5_s_fits_neatly_on_24_gb_cards/
| false | false |
self
| 7 | null |
Qwen did it!
| 350 | 2025-04-28T23:19:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka9ltx/qwen_did_it/
|
josho2001
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka9ltx
| false | null |
t3_1ka9ltx
|
/r/LocalLLaMA/comments/1ka9ltx/qwen_did_it/
| false | false | 350 | null |
||
Does anyone else have any extremely weird benchmarks?
| 8 |
I was recently on a cruise without Internet. It was late. I wasn’t sure if the reception was still open. I really wanted to make sure that I did not miss the sunrise and would set my timer accordingly. I happened to realize that with the amount of data, these LLMs are trained on, in some sense they are almost off-line copies of the Internet. So I tested a few models with prompts in the format: give me your best guess within the minute of the sunrise time on April 20 in Copenhagen. I’ve been trying this on a few models after the cruise for sunrise, sunset, different dates, etc..
I found that closed models like ChatGPT and Gemini do pretty well with guesses within 15 minutes I made sure they didn’t use Internet. Deep seek does poorly with sunset (about 45 minutes off) unless you ask about sunrise first then it’s within 15 minutes. The new best QWEN model does not great with sunset (about 45 minutes off) and even worse when you turn on reasoning (it seriously considered 6:30 PM when the actual sunset was 9:15 PM and used a bunch of nonsense formulas) and is consistently an hour off after reasoning.
But of course, this is not a realistic use case More, just an interesting gauge of its world knowledge so I wanted to ask if any of you have any similar benchmarks
| 2025-04-28T23:23:49 |
https://www.reddit.com/r/LocalLLaMA/comments/1ka9pcw/does_anyone_else_have_any_extremely_weird/
|
Unusual_Guidance2095
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1ka9pcw
| false | null |
t3_1ka9pcw
|
/r/LocalLLaMA/comments/1ka9pcw/does_anyone_else_have_any_extremely_weird/
| false | false |
self
| 8 | null |
Concurrent Test: M3 MAX - Qwen3-30B-A3B [4bit] vs RTX4090 - Qwen3-32B [4bit]
| 25 |
This is a test to compare the token generation speed of the two hardware configurations and new Qwen3 models. Since it is well known that Apple lags behind CUDA in token generation speed, using the MoE model is ideal. For fun, I decided to test both models side by side using the same prompt and parameters, and finally rendering the HTML to compare the quality of the design. I am very impressed with the one-shot design of both models, but Qwen3-32B is truly outstanding.
| 2025-04-28T23:40:37 |
https://v.redd.it/noxzga17wnxe1
|
LocoMod
|
/r/LocalLLaMA/comments/1kaa21l/concurrent_test_m3_max_qwen330ba3b_4bit_vs/
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaa21l
| false |
{'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/noxzga17wnxe1/DASHPlaylist.mpd?a=1748605240%2COWZjNzUwOTAyOTdlZTUzOGE3M2UxNTI3YzU4NDVlNmNlZDgxZjEyYmVlYTIyYzRjOTJlNmU5MzNkYWMyMmY5Ng%3D%3D&v=1&f=sd', 'duration': 137, 'fallback_url': 'https://v.redd.it/noxzga17wnxe1/DASH_1080.mp4?source=fallback', 'has_audio': False, 'height': 1080, 'hls_url': 'https://v.redd.it/noxzga17wnxe1/HLSPlaylist.m3u8?a=1748605240%2CMjc2N2QzZTAzNGNjMTRmNzJjODY2MDNlYzI2ODE5NTYxNjJjMGVhYmEwYmQ1MTI0ODk1ZmI3NjUyZjlhZmYxNA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/noxzga17wnxe1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1880}}
|
t3_1kaa21l
|
/r/LocalLLaMA/comments/1kaa21l/concurrent_test_m3_max_qwen330ba3b_4bit_vs/
| false | false | 25 |
{'enabled': False, 'images': [{'id': 'ZjBvOTZoMTd3bnhlMfHb-as1PWLJytBuE1oXU0NDQiigTnPe3wnkvnlSl9h2', 'resolutions': [{'height': 62, 'url': 'https://external-preview.redd.it/ZjBvOTZoMTd3bnhlMfHb-as1PWLJytBuE1oXU0NDQiigTnPe3wnkvnlSl9h2.png?width=108&crop=smart&format=pjpg&auto=webp&s=f608977fc2ff6af56421df159907856df0f22e87', 'width': 108}, {'height': 124, 'url': 'https://external-preview.redd.it/ZjBvOTZoMTd3bnhlMfHb-as1PWLJytBuE1oXU0NDQiigTnPe3wnkvnlSl9h2.png?width=216&crop=smart&format=pjpg&auto=webp&s=637449fb7933b0fb6ce34ef9a8ff56db0536e3ea', 'width': 216}, {'height': 183, 'url': 'https://external-preview.redd.it/ZjBvOTZoMTd3bnhlMfHb-as1PWLJytBuE1oXU0NDQiigTnPe3wnkvnlSl9h2.png?width=320&crop=smart&format=pjpg&auto=webp&s=8cb94ff94c96d0cd5292c5187bfa28e3e04b12bb', 'width': 320}, {'height': 367, 'url': 'https://external-preview.redd.it/ZjBvOTZoMTd3bnhlMfHb-as1PWLJytBuE1oXU0NDQiigTnPe3wnkvnlSl9h2.png?width=640&crop=smart&format=pjpg&auto=webp&s=6096f4991716f05a4abc70c35ad924fd2bfc85bd', 'width': 640}, {'height': 551, 'url': 'https://external-preview.redd.it/ZjBvOTZoMTd3bnhlMfHb-as1PWLJytBuE1oXU0NDQiigTnPe3wnkvnlSl9h2.png?width=960&crop=smart&format=pjpg&auto=webp&s=22d9214a92b5c72abc25d2edf1e64215ab4f672b', 'width': 960}, {'height': 620, 'url': 'https://external-preview.redd.it/ZjBvOTZoMTd3bnhlMfHb-as1PWLJytBuE1oXU0NDQiigTnPe3wnkvnlSl9h2.png?width=1080&crop=smart&format=pjpg&auto=webp&s=24b9422bd43fd796bd848ceabffc5a8a282ed72e', 'width': 1080}], 'source': {'height': 1986, 'url': 'https://external-preview.redd.it/ZjBvOTZoMTd3bnhlMfHb-as1PWLJytBuE1oXU0NDQiigTnPe3wnkvnlSl9h2.png?format=pjpg&auto=webp&s=9906aaa6ad875788d79b238ba13a9412dcbd1c23', 'width': 3456}, 'variants': {}}]}
|
|
Qwen 3 30B MOE is far better than previous 72B Dense Model
| 49 |
# There is also 32B Dense Model .
CHeck Benchmark ...
|Benchmark|Qwen3-235B-A22B (MoE)|Qwen3-32B (Dense)|OpenAI-o1 (2024-12-17)|Deepseek-R1|Grok 3 Beta (Think)|Gemini2.5-Pro|OpenAI-o3-mini (Medium)|
|:-|:-|:-|:-|:-|:-|:-|:-|
|ArenaHard|95.6|93.8|92.1|93.2|\-|96.4|89.0|
|AIME'24|85.7|81.4|74.3|79.8|83.9|92.0|79.6|
|AIME'25|81.5|72.9|79.2|70.0|77.3|86.7|74.8|
|LiveCodeBench|70.7|65.7|63.9|64.3|70.6|70.4|66.3|
|CodeForces|2056|1977|1891|2029|\-|2001|2036|
|Aider (Pass@2)|61.8|50.2|61.7|56.9|53.3|72.9|53.8|
|LiveBench|77.1|74.9|75.7|71.6|\-|82.4|70.0|
|BFCL|70.8|70.3|67.8|56.9|\-|62.9|64.6|
|MultiIF (8 Langs)|71.9|73.0|48.8|67.7|\-|77.8|48.4|
Full Report:::
[https://qwenlm.github.io/blog/qwen3/](https://qwenlm.github.io/blog/qwen3/)
| 2025-04-28T23:44:23 |
touhidul002
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaa4vh
| false | null |
t3_1kaa4vh
|
/r/LocalLLaMA/comments/1kaa4vh/qwen_3_30b_moe_is_far_better_than_previous_72b/
| false | false | 49 |
{'enabled': True, 'images': [{'id': 'RStXP-fXMDFyVlidS0t-m9S7QzroSChnmFpAvur3bWE', 'resolutions': [{'height': 60, 'url': 'https://preview.redd.it/zmhhrx0yvnxe1.jpeg?width=108&crop=smart&auto=webp&s=ac7ab76f72f4773a8ba7ecc59da79ad2129daa36', 'width': 108}, {'height': 121, 'url': 'https://preview.redd.it/zmhhrx0yvnxe1.jpeg?width=216&crop=smart&auto=webp&s=7933086f3ca5caed5ed247b437f3a7121b8ca43f', 'width': 216}, {'height': 180, 'url': 'https://preview.redd.it/zmhhrx0yvnxe1.jpeg?width=320&crop=smart&auto=webp&s=f8a50a9106f3af68243e47e2803eeb54b45ce2d8', 'width': 320}, {'height': 360, 'url': 'https://preview.redd.it/zmhhrx0yvnxe1.jpeg?width=640&crop=smart&auto=webp&s=4df83b4de32de487ea71eb1d7b6f824a7b834bb2', 'width': 640}, {'height': 540, 'url': 'https://preview.redd.it/zmhhrx0yvnxe1.jpeg?width=960&crop=smart&auto=webp&s=668d8a3b1032d7383ed197aeb6b8dd98f67fbdbf', 'width': 960}, {'height': 607, 'url': 'https://preview.redd.it/zmhhrx0yvnxe1.jpeg?width=1080&crop=smart&auto=webp&s=885162467fae01f79b694f70019358341a50d6b1', 'width': 1080}], 'source': {'height': 1920, 'url': 'https://preview.redd.it/zmhhrx0yvnxe1.jpeg?auto=webp&s=6d52acae75aa8dc6210d87936c80c81d2dd75a67', 'width': 3413}, 'variants': {}}]}
|
||
How does `--cpu-offload-gb` interact with MoE models?
| 2 |
In `vllm` you can do `--cpu-offload-gb`. To load Qwen3-30B-A3B-FP8 this is needed on ~24gb vRAM. My question is given the fact that it's MoE with 3 active params, how much is *actually* in vram at a time? E.g. am I actually going to see a slowdown doing CPU offloading or does this "hack" work in my head
| 2025-04-28T23:47:21 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaa73y/how_does_cpuoffloadgb_interact_with_moe_models/
|
tddammo1
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaa73y
| false | null |
t3_1kaa73y
|
/r/LocalLLaMA/comments/1kaa73y/how_does_cpuoffloadgb_interact_with_moe_models/
| false | false |
self
| 2 | null |
This is 600M parameters??? Yesterday I would have told you this was impossible.
| 402 |
Prompt: "If 5 brog 5 is 1, and 4 brog 2 is 2, what is 3 brog 1?"
thinking content: <think>
Okay, let's see. The problem says that if 5 brog 5 is 1, and 4 brog 2 is 2, then what is 3 brog 1? Hmm, so I need to figure out a pattern or a relationship between the numbers here. Let me start by breaking down the given information.
First, let's parse the question. It mentions "brog" which I assume is a typo or a term that needs to be clarified. Maybe it's supposed to be "bro" or "brog" as in "bro" with a different meaning? Wait, maybe it's a play on words? Like, maybe "brog" stands for something related to numbers or operations. Alternatively, maybe it's a misprint for "bro" which could mean "brog" as in "brog" as in "brog" which is a term used in some contexts, but I'm not sure. Let me think.
Alternatively, perhaps "brog" is a shorthand for multiplication or some other operation. Let me check the examples again. If 5 brog 5 is 1, and 4 brog 2 is 2, then maybe each operation involves multiplying the two numbers? Let me test that. For example, 5 times 5 is 25, which would be 25, but the result given is 1. So that doesn't fit. Alternatively, maybe it's a different operation. Let me think.
Alternatively, maybe it's a combination of numbers. For instance, maybe it's a function f(n, m) where f(n, m) is defined as something. Let's try to think of possible functions. For example, maybe f(n, m) = something like n/m or n + m? Let's check:
If 5 brog 5 is 1, then maybe f(5, 5) = 1. If 4 brog 2 is 2, then f(4, 2) = 2. Let's see if there's a pattern. If we think of 5 brog 5 as 1, which is 5/5 = 1. Then 4 brog 2 is 2, which would be 4/2 = 2. That seems to fit! So maybe the function is f(n, m) = n/m. Then 3 brog 1 would be 3/1 = 3. Therefore, the answer should be 3. But let me check again to make sure there's no other possible interpretation.
Alternatively, maybe there's another operation. For example, maybe it's a combination like n + m, but 5 + 5 is 10, which doesn't match. Or maybe n multiplied by m, but 5\*5 is 25, which doesn't match. Or maybe n divided by m, which would be 1 for the first case and 2 for the second. So that seems consistent. Therefore, if 5 brog 5 is 1, then 3 brog 1 is 3. Therefore, the answer is 3. But let me check if there's another possibility.
Wait, another thought: maybe "brog" is a term related to numbers in a different way. For example, maybe it's a cipher or a code. For example, if "brog" represents something else. But without more context, it's hard to tell. However, given the two examples, multiplying the two numbers gives the result. So 5/5=1, 4/2=2, so 3/1=3. Therefore, I think that's the intended relationship.
Alternatively, maybe it's a function where "brog" represents the number of something. For example, if you have "brog" as in "bro" meaning something else. But again, without knowing the exact meaning, this is speculative.
Alternatively, maybe it's a function where "brog" is a variable. Let me think of a function f(n, m) where n and m are numbers, and f(n, m) is defined in some way. For example, maybe f(n, m) = n - m? Then 5-5=0, which doesn't match. Or f(n, m) = n + m? 5+5=10≠1. Not helpful.
Alternatively, maybe it's a function where the result is the sum of the numbers divided by something. For example, 5 brog 5 is 1. Let me think of fractions. If 5/5=1, then 4/2=2, so 3/1=3. That seems to fit. So maybe the rule is f(n, m) = n/m. Therefore, the answer is 3. Therefore, I think the answer is 3.
\*\*Final Answer\*\*
The value of 3 brog 1 is \\boxed{3}.
</think>
content: To solve the problem, we are given two equations:
\- $ 5 \\text{ brog } 5 = 1 $
\- $ 4 \\text{ brog } 2 = 2 $
We are asked to determine the value of $ 3 \\text{ brog } 1 $.
\---
\### Step 1: Identify the Pattern
We observe that:
\- $ 5 \\text{ brog } 5 = 1 $
\- $ 4 \\text{ brog } 2 = 2 $
If we interpret "brog" as a simple mathematical operation, and assume it is the operation of division, then:
\- $ 5 \\text{ brog } 5 = \\frac{5}{5} = 1 $
\- $ 4 \\text{ brog } 2 = \\frac{4}{2} = 2 $
This pattern suggests that the operation is defined as $ \\frac{n}{m} $, where $ n $ and $ m $ are the numbers in the "brog" expression.
\---
\### Step 2: Apply the Pattern to Find the Result
We are now asked to find the value of $ 3 \\text{ brog } 1 $. Using the same pattern:
$$
3 \\text{ brog } 1 = \\frac{3}{1} = 3
$$
\---
\### Final Answer
$$
\\boxed{3}
$$
| 2025-04-28T23:49:16 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaa8iz/this_is_600m_parameters_yesterday_i_would_have/
|
JohnnyLiverman
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaa8iz
| false | null |
t3_1kaa8iz
|
/r/LocalLLaMA/comments/1kaa8iz/this_is_600m_parameters_yesterday_i_would_have/
| false | false |
self
| 402 | null |
So ... a new qwen 3 32b dense models is even a bit better than 30b moe version
| 30 | 2025-04-28T23:55:13 |
Healthy-Nebula-3603
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaactg
| false | null |
t3_1kaactg
|
/r/LocalLLaMA/comments/1kaactg/so_a_new_qwen_3_32b_dense_models_is_even_a_bit/
| false | false | 30 |
{'enabled': True, 'images': [{'id': 'l2TQawnL6U5_bhrnHrazcoqurVi06JBmy4avXH4lGUg', 'resolutions': [{'height': 125, 'url': 'https://preview.redd.it/mkvvveqyynxe1.png?width=108&crop=smart&auto=webp&s=6005ef46f3910432854a12a944d265fd4099274d', 'width': 108}, {'height': 251, 'url': 'https://preview.redd.it/mkvvveqyynxe1.png?width=216&crop=smart&auto=webp&s=8ca5fb9e49047543250701eb3803a9daee0925f6', 'width': 216}, {'height': 371, 'url': 'https://preview.redd.it/mkvvveqyynxe1.png?width=320&crop=smart&auto=webp&s=25f149784c7cc5b2dfdffd4c453337fbff7adba5', 'width': 320}, {'height': 743, 'url': 'https://preview.redd.it/mkvvveqyynxe1.png?width=640&crop=smart&auto=webp&s=e25449750d834d2998573a6d3234aad4315e8eb1', 'width': 640}], 'source': {'height': 753, 'url': 'https://preview.redd.it/mkvvveqyynxe1.png?auto=webp&s=15bc321a8ef3352a534ca9538f0cacf830876cc2', 'width': 648}, 'variants': {}}]}
|
|||
Qwen 3: What the heck are “Tie Embeddings”?
| 40 |
I thought I had caught up on all the new AI terms out there until I saw “Tie Embeddings” on the Qwen 3 release blog post. Google didn’t really tell me much of anything that I could make any sense of for it. Anyone know what they are and/or why they are important?
| 2025-04-29T00:02:30 |
Porespellar
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaaid5
| false | null |
t3_1kaaid5
|
/r/LocalLLaMA/comments/1kaaid5/qwen_3_what_the_heck_are_tie_embeddings/
| false | false | 40 |
{'enabled': True, 'images': [{'id': 'nLaj9rqtXh5DHAw5tlT4Fq0DYfN3cmeym7PTgdAXFjM', 'resolutions': [{'height': 108, 'url': 'https://preview.redd.it/816fy7jb0oxe1.jpeg?width=108&crop=smart&auto=webp&s=50315adaa9b09fb8f3de811c7648595d6d57944c', 'width': 108}, {'height': 216, 'url': 'https://preview.redd.it/816fy7jb0oxe1.jpeg?width=216&crop=smart&auto=webp&s=22fb124ab56b887f3149dff4e12207451f488f19', 'width': 216}, {'height': 320, 'url': 'https://preview.redd.it/816fy7jb0oxe1.jpeg?width=320&crop=smart&auto=webp&s=9f2b4c3d022e3a98a0624480f625495c09465a5f', 'width': 320}, {'height': 640, 'url': 'https://preview.redd.it/816fy7jb0oxe1.jpeg?width=640&crop=smart&auto=webp&s=c95a7c23fda5c6da6b008d2b00554f9bd6b9ed0d', 'width': 640}, {'height': 960, 'url': 'https://preview.redd.it/816fy7jb0oxe1.jpeg?width=960&crop=smart&auto=webp&s=cd49fa2995161e48852949fb593b52d214bb0e7d', 'width': 960}, {'height': 1080, 'url': 'https://preview.redd.it/816fy7jb0oxe1.jpeg?width=1080&crop=smart&auto=webp&s=c677c1bf2fa56a0f709c879a7e0ce611a95fa076', 'width': 1080}], 'source': {'height': 1125, 'url': 'https://preview.redd.it/816fy7jb0oxe1.jpeg?auto=webp&s=eb7b8dbb67da6ad721ca9e805831ca3b05e44d06', 'width': 1125}, 'variants': {}}]}
|
||
Can Qwen3-235B-A22B run efficiently on my hardware(256gb ram+quad 3090s ) with vLLM?
| 3 |
I've been reading about Qwen3-30B-A3B and understand that it only activates 3B parameters at runtime while the total model is 30B (which explains why it can run at 20 tps even on a 4GB GPU
link: [https://www.reddit.com/r/LocalLLaMA/comments/1ka8n18/qwen330ba3b\_is\_magic](https://www.reddit.com/r/LocalLLaMA/comments/1ka8n18/qwen330ba3b_is_magic) ).
I'm interested in running the larger **Qwen3-235B-A22B-FP8** model using the same MoE (Mixture of Experts) principle where only 22B parameters are activated during inference.
My current hardware setup:
* 256GB system RAM
* Intel 10900X CPU
* 4× RTX 3090 GPUs in quad configuration
I'm wondering if vLLM can efficiently serve this model by:
1. Loading only the required experts into GPU memory (the active 22B parameters)
2. Keeping the rest of the model in system RAM
3. Dynamically swapping experts as needed during inference
Has anyone tried running this specific configuration? What kind of performance could I expect? Any specific settings I should use to optimize for this hardware?
| 2025-04-29T00:14:00 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaaqti/can_qwen3235ba22b_run_efficiently_on_my/
|
Acceptable-State-271
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaaqti
| false | null |
t3_1kaaqti
|
/r/LocalLLaMA/comments/1kaaqti/can_qwen3235ba22b_run_efficiently_on_my/
| false | false |
self
| 3 | null |
Fastest inference on Mac: MLX, llama.cpp, vLLM, exLlamav2, sglang?
| 2 |
I'm trying to do [batch inference for long document QA](https://www.reddit.com/r/LocalLLaMA/comments/1k5pwq5/fastestbest_way_for_local_llms_to_answer_many/), and my Mac is doing it really slowly on llama.cpp: about 4 tok/s for Mistral-Nemo-Instruct-2407-Q4\_K\_M.gguf with 36gb RAM, which takes an hour per patient.
I run llama.cpp with \`llama-server -m Mistral-Nemo-Instruct-2407-Q4\_K\_M.gguf -c 16384 --port 8081 -ngl -1 -np 2\` and I get:
\`\`\`
prompt eval time = 24470.27 ms / 3334 tokens ( 7.34 ms per token, 136.25 tokens per second)
eval time = 82158.50 ms / 383 tokens ( 214.51 ms per token, 4.66 tokens per second)
total time = 106628.78 ms / 3717 tokens
\`\`\`
I'm not sure if other frameworks like MLX/vLLM/exLlamaV2 are faster, but the speed is a big problem in my pipeline.
The vLLM documentation suggests that it only works well on Linux and that compiling it for Mac makes it CPU only, which doesn't sound very promising.
| 2025-04-29T00:16:36 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaasp4/fastest_inference_on_mac_mlx_llamacpp_vllm/
|
Amazydayzee
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaasp4
| false | null |
t3_1kaasp4
|
/r/LocalLLaMA/comments/1kaasp4/fastest_inference_on_mac_mlx_llamacpp_vllm/
| false | false |
self
| 2 | null |
Bug in Unsloth Qwen3 GGUF chat template?
| 18 |
When I load unsloth Qwen3 GGUFs (checked 32B and 4B), I get this error:
```
common_chat_templates_init: failed to parse chat template (defaulting to chatml): Expected value expression at row 18, column 30:
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
{%- for message in messages[::-1] %}
^
```
I think maybe they meant to write:
```
{%- for message in messages | reverse %}
```
The failure to load the template means llama.cpp defaults to chatml template.
| 2025-04-29T00:40:17 |
https://www.reddit.com/r/LocalLLaMA/comments/1kab9po/bug_in_unsloth_qwen3_gguf_chat_template/
|
DeltaSqueezer
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kab9po
| false | null |
t3_1kab9po
|
/r/LocalLLaMA/comments/1kab9po/bug_in_unsloth_qwen3_gguf_chat_template/
| false | false |
self
| 18 | null |
Why all thinking local LLM's keep doing this for me? What setting do I need to change or what system prompt should I have?
| 3 |
Tried running the same model online, and it was perfect, didn't even go into thinking mode, just gave me correct answers. Locally, the same model does this for some reason.
| 2025-04-29T00:48:40 |
https://v.redd.it/kb28latt7oxe1
|
Leoxooo
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabfqk
| false |
{'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/kb28latt7oxe1/DASHPlaylist.mpd?a=1748479737%2CODQ1Mzk1MjQyYjJkODZmZTY3MTk5YzM0MmJiMjdmNDFkNWUzMzRhMjAyZjEzZTUyZGFjODQ5YWVjYWYxNWM3Nw%3D%3D&v=1&f=sd', 'duration': 35, 'fallback_url': 'https://v.redd.it/kb28latt7oxe1/DASH_1080.mp4?source=fallback', 'has_audio': False, 'height': 1080, 'hls_url': 'https://v.redd.it/kb28latt7oxe1/HLSPlaylist.m3u8?a=1748479737%2CYzI2MmQxMGZlNTY0NjI5ZjU4M2VjZTI1Yjk2NjhkMDY1ZTg5ZmE3OWIxMjI2NDAwNzg4NzY3NzkyNmQwNDI3MA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/kb28latt7oxe1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1224}}
|
t3_1kabfqk
|
/r/LocalLLaMA/comments/1kabfqk/why_all_thinking_local_llms_keep_doing_this_for/
| false | false | 3 |
{'enabled': False, 'images': [{'id': 'bmloZ3M5dHQ3b3hlMT_tMAY4tU-JuT4iGuurBm9ODoPZV_HXsFMoU4kYvF3R', 'resolutions': [{'height': 95, 'url': 'https://external-preview.redd.it/bmloZ3M5dHQ3b3hlMT_tMAY4tU-JuT4iGuurBm9ODoPZV_HXsFMoU4kYvF3R.png?width=108&crop=smart&format=pjpg&auto=webp&s=7649d61db76134656692de1e4f44574c66cc15d2', 'width': 108}, {'height': 190, 'url': 'https://external-preview.redd.it/bmloZ3M5dHQ3b3hlMT_tMAY4tU-JuT4iGuurBm9ODoPZV_HXsFMoU4kYvF3R.png?width=216&crop=smart&format=pjpg&auto=webp&s=aa5a14c59b2760633f769cfd41c2f01206fd0ce5', 'width': 216}, {'height': 282, 'url': 'https://external-preview.redd.it/bmloZ3M5dHQ3b3hlMT_tMAY4tU-JuT4iGuurBm9ODoPZV_HXsFMoU4kYvF3R.png?width=320&crop=smart&format=pjpg&auto=webp&s=fc400d092cf7c089f5efd301d6e211a62a2c6249', 'width': 320}, {'height': 564, 'url': 'https://external-preview.redd.it/bmloZ3M5dHQ3b3hlMT_tMAY4tU-JuT4iGuurBm9ODoPZV_HXsFMoU4kYvF3R.png?width=640&crop=smart&format=pjpg&auto=webp&s=d1c1381092e3c273bc2b08d1136366a8c7c61a4a', 'width': 640}, {'height': 847, 'url': 'https://external-preview.redd.it/bmloZ3M5dHQ3b3hlMT_tMAY4tU-JuT4iGuurBm9ODoPZV_HXsFMoU4kYvF3R.png?width=960&crop=smart&format=pjpg&auto=webp&s=11bde59b22e32fb1fb01cd3396191767a4dabdc5', 'width': 960}, {'height': 953, 'url': 'https://external-preview.redd.it/bmloZ3M5dHQ3b3hlMT_tMAY4tU-JuT4iGuurBm9ODoPZV_HXsFMoU4kYvF3R.png?width=1080&crop=smart&format=pjpg&auto=webp&s=6afbc04a94ce97d340d25db025f935a6ba5c7c51', 'width': 1080}], 'source': {'height': 1488, 'url': 'https://external-preview.redd.it/bmloZ3M5dHQ3b3hlMT_tMAY4tU-JuT4iGuurBm9ODoPZV_HXsFMoU4kYvF3R.png?format=pjpg&auto=webp&s=88968ad4e613c5156cb38504c2fb019aa2623046', 'width': 1686}, 'variants': {}}]}
|
|
VULKAN is faster tan CUDA currently with LLAMACPP! 62.2 T/S vs 77.5 t/s
| 117 | 2025-04-29T00:53:40 |
https://www.reddit.com/r/LocalLLaMA/comments/1kabje8/vulkan_is_faster_tan_cuda_currently_with_llamacpp/
|
Healthy-Nebula-3603
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabje8
| false | null |
t3_1kabje8
|
/r/LocalLLaMA/comments/1kabje8/vulkan_is_faster_tan_cuda_currently_with_llamacpp/
| false | false | 117 | null |
||
RE: Qwen 3
| 1 |
[removed]
| 2025-04-29T00:56:48 |
https://www.reddit.com/r/LocalLLaMA/comments/1kabll5/re_qwen_3/
|
MerePotato
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabll5
| false | null |
t3_1kabll5
|
/r/LocalLLaMA/comments/1kabll5/re_qwen_3/
| false | false |
self
| 1 | null |
Regarding Qwen 3
| 1 |
[removed]
| 2025-04-29T00:58:38 |
https://www.reddit.com/r/LocalLLaMA/comments/1kabmw4/regarding_qwen_3/
|
MerePotato
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabmw4
| false | null |
t3_1kabmw4
|
/r/LocalLLaMA/comments/1kabmw4/regarding_qwen_3/
| false | false |
self
| 1 | null |
Prototype Synthetic RP Dataset
| 5 |
This has been in the works for a while now, and I was hoping to get a little feedback. Right now, I'm only at about 20 turns for a little over 9,000 character cards. I wanted to get a little more feedback before continuing.
You can read the dataset card for more info. I tried to make it funny. But TLDR, I took a few thousand chub/janitorai/whatever cards, generated some synthetic "improved cards" and mixed them all together. Then I used Llama Maverick to generate the first few messages of the conversation. Once that's done, I switched to Deepseek chat. People really seem to hate on Maverick, but it seems less censored by default, and giving Deepseek Maverick messages to start with seems to really help with the Deepseek "unhinged factor". And Deepseek refuses way less once there are already non refusal examples messages. I also did a psychoanalysis pass on each character card to help give the synthetic "human user" more personality to complement the character card, helping indicate that kind of roleplay the person who chose that card might want. Eventually I want to use this pipeline to generate some real crazy "exotic alignment" datasets, but I need to get the basics down first.
I built a script for creating multi turn data to help make this dataset, I'll probably release that too once I make it look a little bit less like code spaghetti. I still need to clean this data up most likely and run some more validation. But I'm interested if anyone has ideas for how I could make this better. Eventually I want a huge long context roleplay dataset I could train a much smaller model on, using all open source data. I'm curious what people think of this idea.
Good start? Or start over?
| 2025-04-29T00:59:10 |
https://huggingface.co/datasets/openerotica/long-roleplay-v0.1
|
Scam_Altman
|
huggingface.co
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabn8w
| false | null |
t3_1kabn8w
|
/r/LocalLLaMA/comments/1kabn8w/prototype_synthetic_rp_dataset/
| false | false | 5 |
{'enabled': False, 'images': [{'id': 'm3suvXVCvL5DzE8Jv-cwnAbsdEdN0oMhILOJ9tBX36w', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/NEOsR7D67c5yihn8kPe8xfT1lH0ur6s5BNpV9Jt_ANU.jpg?width=108&crop=smart&auto=webp&s=3e1caa5b1bb5be37fe3203ecab497e338058bd5f', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/NEOsR7D67c5yihn8kPe8xfT1lH0ur6s5BNpV9Jt_ANU.jpg?width=216&crop=smart&auto=webp&s=4dba6061a3591aa5d4f75af8f2cb7e72889c9ec0', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/NEOsR7D67c5yihn8kPe8xfT1lH0ur6s5BNpV9Jt_ANU.jpg?width=320&crop=smart&auto=webp&s=23e708f3e5559fc1d04a419ee79e9b997c3bb4ef', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/NEOsR7D67c5yihn8kPe8xfT1lH0ur6s5BNpV9Jt_ANU.jpg?width=640&crop=smart&auto=webp&s=9e0477ee9d4ccace44424d05f1e8886ac93db281', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/NEOsR7D67c5yihn8kPe8xfT1lH0ur6s5BNpV9Jt_ANU.jpg?width=960&crop=smart&auto=webp&s=51ffb488482e05e19cebe5765aba1bedc2e3334e', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/NEOsR7D67c5yihn8kPe8xfT1lH0ur6s5BNpV9Jt_ANU.jpg?width=1080&crop=smart&auto=webp&s=912c1ff96ec61f03abfe9b793f78aaf65cbd0f61', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/NEOsR7D67c5yihn8kPe8xfT1lH0ur6s5BNpV9Jt_ANU.jpg?auto=webp&s=c1330e4e6e8ad545e6cb17f7412855ba56d868d2', 'width': 1200}, 'variants': {}}]}
|
|
Is Qwen3 doing benchmaxxing?
| 65 |
Very good benchmarks scores. But some early indication suggests that it's not as good as the benchmarks suggests.
What are your findings?
| 2025-04-29T00:59:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1kabnca/is_qwen3_doing_benchmaxxing/
|
EasternBeyond
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabnca
| false | null |
t3_1kabnca
|
/r/LocalLLaMA/comments/1kabnca/is_qwen3_doing_benchmaxxing/
| false | false |
self
| 65 | null |
Regarding Qwen 3
| 1 |
[removed]
| 2025-04-29T01:03:50 |
https://www.reddit.com/r/LocalLLaMA/comments/1kabqm6/regarding_qwen_3/
|
MerePotato
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabqm6
| false | null |
t3_1kabqm6
|
/r/LocalLLaMA/comments/1kabqm6/regarding_qwen_3/
| false | false |
self
| 1 | null |
Do not use Qwen3-30-A3B with quantized kv cache
| 1 |
Just tested A3B model in ollama, with Q8 cache, and it has the same issue of Granite 2B and small Gemma3 models(4B and 12B):
It's really slow then it should be and don't utilize GPU well, my 4090 is running at 100w when running this model
The issue is Quantized kV cache:
https://github.com/ggml-org/llama.cpp/issues/12352#issuecomment-2727452955
It seems more of a hardware issue then software one, but since this is a new model, idk, maybe it could be fixed?
| 2025-04-29T01:05:08 |
https://www.reddit.com/r/LocalLLaMA/comments/1kabrhm/do_not_use_qwen330a3b_with_quantized_kv_cache/
|
AaronFeng47
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabrhm
| false | null |
t3_1kabrhm
|
/r/LocalLLaMA/comments/1kabrhm/do_not_use_qwen330a3b_with_quantized_kv_cache/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'hONGl-A2cszRrFphsgv0NNfiF8_TWh29BIKW1lV5YbU', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/u6bYa7kOT8YBTX95hisG538axVlWtaav7Mni51AlN84.jpg?width=108&crop=smart&auto=webp&s=4ff3da96e59f10b72f2f81243874fc4405b0b744', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/u6bYa7kOT8YBTX95hisG538axVlWtaav7Mni51AlN84.jpg?width=216&crop=smart&auto=webp&s=449a2e5fbc08fb839cac2c74518675c45ee01d28', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/u6bYa7kOT8YBTX95hisG538axVlWtaav7Mni51AlN84.jpg?width=320&crop=smart&auto=webp&s=18c5d9781456897232a91747e354ce1c81a89c55', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/u6bYa7kOT8YBTX95hisG538axVlWtaav7Mni51AlN84.jpg?width=640&crop=smart&auto=webp&s=c6c029de52774c7a36ea370a81da3b844815dd3f', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/u6bYa7kOT8YBTX95hisG538axVlWtaav7Mni51AlN84.jpg?width=960&crop=smart&auto=webp&s=02cd831115e0b74133319446e3d8863bafdab367', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/u6bYa7kOT8YBTX95hisG538axVlWtaav7Mni51AlN84.jpg?width=1080&crop=smart&auto=webp&s=24189568de0b243d740bde489c978af790506d0f', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/u6bYa7kOT8YBTX95hisG538axVlWtaav7Mni51AlN84.jpg?auto=webp&s=3fc355a0ecb27ff3a76cbe3486775e1b044d9e51', 'width': 1200}, 'variants': {}}]}
|
Scaling Peer-To-Peer Decentralized Inference
| 3 |
We are excited to share a preview of our peer-to-peer decentralized inference stack — engineered for consumer GPUs and the 100ms latencies of the public internet—plus a research roadmap that scales it into a **planetary-scale inference engine**.
At Prime Intellect, we’re building towards an open and decentralized AGI future—one where anyone with consumer-grade hardware and a network connection can meaningfully contribute to and benefit from AGI. This means designing for the real world: heterogeneous GPUs, public internet latency, and unreliable but abundant FLOPs. With the rise of reinforcement learning for reasoning models like DeepSeek R1, inference has moved to center stage, and is now a core component of the entire AI stack:
* **Training**: Generate rollouts during reinforcement learning (e.g. INTELLECT-2)
* **Distillation**: Creating synthetic data at scale (e.g. SYNTHETIC-1)
* **Evaluation**: Benchmarking model performance and safety
That’s why our next step is decentralizing inference itself.
| 2025-04-29T01:09:36 |
https://www.primeintellect.ai/blog/inference
|
primeintellect_ai
|
primeintellect.ai
| 1970-01-01T00:00:00 | 0 |
{}
|
1kabujt
| false | null |
t3_1kabujt
|
/r/LocalLLaMA/comments/1kabujt/scaling_peertopeer_decentralized_inference/
| false | false | 3 |
{'enabled': False, 'images': [{'id': 'y4du1q97bWGKbEDXs9G6MNRIWSQ1anbLEl4Maitm7wI', 'resolutions': [{'height': 60, 'url': 'https://external-preview.redd.it/unvWMDY_pgvnh7_49n7mRD0K5iOTu-T5GMvbLvSs2y4.jpg?width=108&crop=smart&auto=webp&s=a074e153fa0ccbeb9d337797459be79f2423277f', 'width': 108}, {'height': 121, 'url': 'https://external-preview.redd.it/unvWMDY_pgvnh7_49n7mRD0K5iOTu-T5GMvbLvSs2y4.jpg?width=216&crop=smart&auto=webp&s=198789afeb392120efde17a06170d5b5a92f3446', 'width': 216}, {'height': 180, 'url': 'https://external-preview.redd.it/unvWMDY_pgvnh7_49n7mRD0K5iOTu-T5GMvbLvSs2y4.jpg?width=320&crop=smart&auto=webp&s=c0463b367fa970370af4b8fc511514fd38cf40a5', 'width': 320}, {'height': 360, 'url': 'https://external-preview.redd.it/unvWMDY_pgvnh7_49n7mRD0K5iOTu-T5GMvbLvSs2y4.jpg?width=640&crop=smart&auto=webp&s=ed7a480b0720935cdadc53a290777050848289bd', 'width': 640}, {'height': 540, 'url': 'https://external-preview.redd.it/unvWMDY_pgvnh7_49n7mRD0K5iOTu-T5GMvbLvSs2y4.jpg?width=960&crop=smart&auto=webp&s=81d2e5fecd7c3ef46228cf434e9899810bb866ce', 'width': 960}, {'height': 608, 'url': 'https://external-preview.redd.it/unvWMDY_pgvnh7_49n7mRD0K5iOTu-T5GMvbLvSs2y4.jpg?width=1080&crop=smart&auto=webp&s=fe0a91af06099064864f3582980f501b30d5ef54', 'width': 1080}], 'source': {'height': 2913, 'url': 'https://external-preview.redd.it/unvWMDY_pgvnh7_49n7mRD0K5iOTu-T5GMvbLvSs2y4.jpg?auto=webp&s=afce00baf6e525d10ce9c6a9f9a2d5aba0d22350', 'width': 5170}, 'variants': {}}]}
|
|
Qwen3 token budget
| 7 |
Hats off to the Qwen team for such a well-planned release with day 0 support, unlike, ironically, llama.
Anyways, I read on their blog that token budgets are a thing, similar to (I think) claude 3.7 sonnet. They show some graphs with performance increases with longer budgets.
Anyone know how to actually set these? I would assume token cutoff is definetly not it, as that would cut off the response.
Did they just use token cutoff and in the next prompt tell the model to provide a final answer?
| 2025-04-29T01:18:21 |
https://www.reddit.com/r/LocalLLaMA/comments/1kac0qh/qwen3_token_budget/
|
dp3471
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kac0qh
| false | null |
t3_1kac0qh
|
/r/LocalLLaMA/comments/1kac0qh/qwen3_token_budget/
| false | false |
self
| 7 | null |
Qwen3: How to Run & Fine-tune | Unsloth
| 11 |
Non-Thinking Mode Settings:
Temperature = 0.7
Min_P = 0.0 (optional, but 0.01 works well, llama.cpp default is 0.1)
Top_P = 0.8
TopK = 20
Thinking Mode Settings:
Temperature = 0.8
Min_P = 0.0
Top_P = 0.95
TopK = 20
https://docs.unsloth.ai/basics/qwen3-how-to-run-and-fine-tune
| 2025-04-29T01:19:26 |
https://www.reddit.com/r/LocalLLaMA/comments/1kac1gi/qwen3_how_to_run_finetune_unsloth/
|
slypheed
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kac1gi
| false | null |
t3_1kac1gi
|
/r/LocalLLaMA/comments/1kac1gi/qwen3_how_to_run_finetune_unsloth/
| false | false |
self
| 11 |
{'enabled': False, 'images': [{'id': 'ZmadbtMLxXXHFKwJkCjeTUDuX5sS57sYwkHR8IIGo6Y', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/fxYCW6fqdbJ5RWjh_x1fsIyj0ZtZFx8MOAvXVxIw2PE.jpg?width=108&crop=smart&auto=webp&s=1ef4773905a7285d6ca9d2707252ecf3322ec746', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/fxYCW6fqdbJ5RWjh_x1fsIyj0ZtZFx8MOAvXVxIw2PE.jpg?width=216&crop=smart&auto=webp&s=6555cce3e1543ec541933b9a1ea746f3da79448a', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/fxYCW6fqdbJ5RWjh_x1fsIyj0ZtZFx8MOAvXVxIw2PE.jpg?width=320&crop=smart&auto=webp&s=346b61e1006578bd8c7c90ff8b45496164cd4933', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/fxYCW6fqdbJ5RWjh_x1fsIyj0ZtZFx8MOAvXVxIw2PE.jpg?width=640&crop=smart&auto=webp&s=2e74df95b54af72feafa558281ef5e11bc4e8a7c', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/fxYCW6fqdbJ5RWjh_x1fsIyj0ZtZFx8MOAvXVxIw2PE.jpg?width=960&crop=smart&auto=webp&s=8d3ac1cc3775d1b7217345a94a6e9f18f0ba2092', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/fxYCW6fqdbJ5RWjh_x1fsIyj0ZtZFx8MOAvXVxIw2PE.jpg?width=1080&crop=smart&auto=webp&s=57e2a43db692dc32eecd433adfbae429f9bca7fd', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/fxYCW6fqdbJ5RWjh_x1fsIyj0ZtZFx8MOAvXVxIw2PE.jpg?auto=webp&s=2704eae76891f7897192cd5a7236096d2b9f8a5f', 'width': 1200}, 'variants': {}}]}
|
Qwen3 32B: How many r's are there in the word strawberrrrrrrrrrry?
| 0 |
<think>
Okay, let's see. The user is asking how many 'r's are in the word "strawberrrrrrrrrrry". Hmm, first I need to make sure I have the correct spelling of the word. Let me write it out: S-T-R-A-W-B-E-R-R-R-R-R-R-R-R-R-R-Y. Wait, that looks like a lot of 'r's towards the end. Maybe the user is testing my attention to detail here. Alright, let's break it down. The word is "strawberrrrrrrrrrry". Let me count each letter step by step. Starting from the beginning: S - that's the first letter, not an 'r'. T - second letter, still no 'r'. R - third letter, that's one 'r' so far. A - fourth, not an 'r'. W - fifth, nope. B - sixth, nope. E - seventh, nope. Now we get to the next part. The next letters after 'berry'... Wait, normally "strawberry" has two 'r's in "berry". But here, the user has written "berrrrrrrrrrry", which seems like a lot more 'r's. Let me check again. Original word: S-T-R-A-W-B-E-R followed by multiple 'r's and ending with Y. Let me count each 'r' in the entire word. Breaking it down:
* The first 'r' is in the third position: S-T-R...
* Then after the 'berry' part: B-E-R-R... So the 'berry' part normally has two 'r's, but in this case, it's spelled with a bunch of 'r's. Let me count them all. Let me write out the letters one by one: S T R A W B E R R R R R R R R R R R Y Wait, let me verify the positions again. The word is "strawberrrrrrrrrrry". Let's count each letter: S T R A W B E R R R R R R R R R R R Y So from position 3: one 'r', then starting at position 8, there are 10 'r's in a row? Let me count again. After the 'E' (position 7), the next letters are R (8), R (9), R (10), R (11), R (12), R (13), R (14), R (15), R (16), R (17), R (18). Wait, that's 11 'r's? Let me check again: After E (position 7), the letters are:
1. R
2. R
3. R
4. R
5. R
6. R
7. R
8. R
9. R
10. R
11. R
12. Y So from position 8 to 18, that's 11 'r's. Then the last letter is Y. So total 'r's would be the one at position 3 plus the 11 from positions 8-18. That's 1 + 11 = 12 'r's. Wait, but the original word is "strawberrrrrrrrrrry". Let me count the 'r's in the string "berrrrrrrrrrry". The 'berry' part here is written as 'berrrrrrrrrrry', which is B-E followed by 11 'r's and then Y. So in that part, the 'r's are 11. Then the first 'r' is in the 'straw' part: S-T-R-A-W... So that's one 'r'. So total is 1 + 11 = 12 'r's. But let me double-check by writing out the entire word with letters: S T R A W B E R R R R R R R R R R R Y Breaking it down:
* S
* T
* R (1)
* A
* W
* B
* E
* R (2)
* R (3)
* R (4)
* R (5)
* R (6)
* R (7)
* R (8)
* R (9)
* R (10)
* R (11)
* R (12)
* Y Yes, that's 12 'r's. The first 'r' is in the third position, then from position 8 to 18, there are 11 'r's. So total of 12 'r's. I think that's correct. Let me count again to be sure. The first 'r' is one. Then after the 'E', the next letters are 11 'r's. So 1 + 11 = 12. Yeah, that seems right.
</think>
The word "strawberrrrrrrrrrry" contains **12** instances of the letter 'r'. **Breakdown:**
* The first 'r' appears in the third position: **S-T-R**...
* After the 'berry' part, there are 11 consecutive 'r's: **B-E-R-R-R-R-R-R-R-R-R-R-R**... **Total 'r's:** 1 (initial) + 11 (in the 'berry' part) = **12**.
| 2025-04-29T01:25:45 |
https://www.reddit.com/r/LocalLLaMA/comments/1kac5wc/qwen3_32b_how_many_rs_are_there_in_the_word/
|
NNN_Throwaway2
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kac5wc
| false | null |
t3_1kac5wc
|
/r/LocalLLaMA/comments/1kac5wc/qwen3_32b_how_many_rs_are_there_in_the_word/
| false | false |
self
| 0 | null |
Unsloth is uploading 128K context Qwen3 GGUFs
| 72 |
[https://huggingface.co/models?search=unsloth%20qwen3%20128k](https://huggingface.co/models?search=unsloth%20qwen3%20128k)
https://preview.redd.it/36gv4a9lgoxe1.png?width=631&format=png&auto=webp&s=3d8a861695b959775b32aadc027ca0195d3e81c3
https://preview.redd.it/p937wanlgoxe1.png?width=386&format=png&auto=webp&s=201e7d0a664a2db75938da7482f14e025d2519b6
Plus their [Qwen3-30B-A3B-GGUF](https://huggingface.co/unsloth/Qwen3-30B-A3B-GGUF) might have some bugs:
https://preview.redd.it/s3dtp56ugoxe1.png?width=561&format=png&auto=webp&s=a82fc690b5746f668b8b00fb4520004c710da7dd
| 2025-04-29T01:35:12 |
https://www.reddit.com/r/LocalLLaMA/comments/1kacch6/unsloth_is_uploading_128k_context_qwen3_ggufs/
|
AaronFeng47
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kacch6
| false | null |
t3_1kacch6
|
/r/LocalLLaMA/comments/1kacch6/unsloth_is_uploading_128k_context_qwen3_ggufs/
| false | false | 72 |
{'enabled': False, 'images': [{'id': 'An0iJLapq-5CUQQlm3lWegevVWf7wlANjmn1iOwCTqk', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/BOaSYNg6lhlngBhuDS68WpIBibLf88Q_KzjZVrFpgEc.jpg?width=108&crop=smart&auto=webp&s=2c0b032bdc9d0820b318f57def3af620afe60ee8', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/BOaSYNg6lhlngBhuDS68WpIBibLf88Q_KzjZVrFpgEc.jpg?width=216&crop=smart&auto=webp&s=7b29327d787489e6d4f61726ba9d10a09ed099d9', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/BOaSYNg6lhlngBhuDS68WpIBibLf88Q_KzjZVrFpgEc.jpg?width=320&crop=smart&auto=webp&s=9f1b5bed20b4b058b596c2a430a47d3b9c857e03', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/BOaSYNg6lhlngBhuDS68WpIBibLf88Q_KzjZVrFpgEc.jpg?width=640&crop=smart&auto=webp&s=7b47505d7a8ebd834ca805c293d16277b5772c12', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/BOaSYNg6lhlngBhuDS68WpIBibLf88Q_KzjZVrFpgEc.jpg?width=960&crop=smart&auto=webp&s=c7be2b4b0ad69f9ff176d6a0027458c22a63a5f0', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/BOaSYNg6lhlngBhuDS68WpIBibLf88Q_KzjZVrFpgEc.jpg?width=1080&crop=smart&auto=webp&s=dea3a5ccadcdb95c05dca40d482f50c976b88233', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/BOaSYNg6lhlngBhuDS68WpIBibLf88Q_KzjZVrFpgEc.jpg?auto=webp&s=6e3e4780238d40a2755c2289e7e3d722eeb8ea30', 'width': 1200}, 'variants': {}}]}
|
|
Run Qwen3 (0.6B) 100% locally in your browser on WebGPU w/ Transformers.js
| 135 | 2025-04-29T01:48:54 |
https://v.redd.it/wuflehwxioxe1
|
xenovatech
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaclyw
| false |
{'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/wuflehwxioxe1/DASHPlaylist.mpd?a=1748483350%2CNTZhMGFmNDA2NzFmNTBiYWVlYjNiZGMzZWZhY2I1ZWYzMjMxYTg4YzUxYTUyMmI0YzQ4OTc3NWI1OTMxZDA0NA%3D%3D&v=1&f=sd', 'duration': 24, 'fallback_url': 'https://v.redd.it/wuflehwxioxe1/DASH_1080.mp4?source=fallback', 'has_audio': False, 'height': 1082, 'hls_url': 'https://v.redd.it/wuflehwxioxe1/HLSPlaylist.m3u8?a=1748483350%2COGY0N2ZkMzQ1ZWRmNTFmZjc1ZDQ0NjU2ODE1YzMxYmZkNTk5ZDJiMjVjZGY3NjhhZDc2NmZkMzZmZTZlZTg1NA%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/wuflehwxioxe1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1080}}
|
t3_1kaclyw
|
/r/LocalLLaMA/comments/1kaclyw/run_qwen3_06b_100_locally_in_your_browser_on/
| false | false | 135 |
{'enabled': False, 'images': [{'id': 'a3VnbHRmd3hpb3hlMbRgvsKi_IaAGC-viwKm_brdcymj4zDazyH2qxIAGtXM', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/a3VnbHRmd3hpb3hlMbRgvsKi_IaAGC-viwKm_brdcymj4zDazyH2qxIAGtXM.png?width=108&crop=smart&format=pjpg&auto=webp&s=69209da8f5f8a4fbbd8c7274ee73d49ed7ed7742', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/a3VnbHRmd3hpb3hlMbRgvsKi_IaAGC-viwKm_brdcymj4zDazyH2qxIAGtXM.png?width=216&crop=smart&format=pjpg&auto=webp&s=b7b66a05aaa67edb89221c7dcd43e5625c722164', 'width': 216}, {'height': 320, 'url': 'https://external-preview.redd.it/a3VnbHRmd3hpb3hlMbRgvsKi_IaAGC-viwKm_brdcymj4zDazyH2qxIAGtXM.png?width=320&crop=smart&format=pjpg&auto=webp&s=258f13946416f6ee7a8aa713da078270ae46e085', 'width': 320}, {'height': 640, 'url': 'https://external-preview.redd.it/a3VnbHRmd3hpb3hlMbRgvsKi_IaAGC-viwKm_brdcymj4zDazyH2qxIAGtXM.png?width=640&crop=smart&format=pjpg&auto=webp&s=c4ecae12e92d91cbba06916c1c1d760392d52b69', 'width': 640}, {'height': 961, 'url': 'https://external-preview.redd.it/a3VnbHRmd3hpb3hlMbRgvsKi_IaAGC-viwKm_brdcymj4zDazyH2qxIAGtXM.png?width=960&crop=smart&format=pjpg&auto=webp&s=1c9b0a5e31d983d62dd3446ac2aac84179f7a948', 'width': 960}, {'height': 1081, 'url': 'https://external-preview.redd.it/a3VnbHRmd3hpb3hlMbRgvsKi_IaAGC-viwKm_brdcymj4zDazyH2qxIAGtXM.png?width=1080&crop=smart&format=pjpg&auto=webp&s=d89bb427428ce8138c23e2d6fb0db52a6a2a64b0', 'width': 1080}], 'source': {'height': 1602, 'url': 'https://external-preview.redd.it/a3VnbHRmd3hpb3hlMbRgvsKi_IaAGC-viwKm_brdcymj4zDazyH2qxIAGtXM.png?format=pjpg&auto=webp&s=37bdd34e692071616244434d832b17e2e55ad92e', 'width': 1600}, 'variants': {}}]}
|
||
Are most improvements in models from continuous fine tuning rather than architecture changes?
| 6 |
Most models like Qwen2.5 or Llama3.3 seem to just be scaled up versions of GPT 2 architecture, following the decoder block diagram of the “attention is all you need” paper. I noticed the activation functions changed, and maybe the residuals swapped places with the normalization for some (?) but everything else seems to be relatively similar. Does that mean the full potential and limits of the decoder only model have not been reached yet?
I know mixture of experts and latent attention exist, but many decoder only’s when scaled up preform similarly.
| 2025-04-29T02:02:24 |
https://www.reddit.com/r/LocalLLaMA/comments/1kacvdj/are_most_improvements_in_models_from_continuous/
|
Ok-Cicada-5207
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kacvdj
| false | null |
t3_1kacvdj
|
/r/LocalLLaMA/comments/1kacvdj/are_most_improvements_in_models_from_continuous/
| false | false |
self
| 6 | null |
Qwen3 Censorship
| 0 |
Any Qwen3 uncensored models yet?
| 2025-04-29T02:03:43 |
https://www.reddit.com/r/LocalLLaMA/comments/1kacwah/qwen3_censorship/
|
getSAT
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kacwah
| false | null |
t3_1kacwah
|
/r/LocalLLaMA/comments/1kacwah/qwen3_censorship/
| false | false |
self
| 0 | null |
Qwen 3 wants to respond in Chinese, even when not in prompt.
| 16 |
For short basic prompts I seem to be triggering responses in Chinese often, where it says "Also, need to make sure the response is in Chinese, as per the user's preference. Let me check the previous interactions to confirm the language. Yes, previous responses are in Chinese. So I'll structure the answer to be honest yet supportive, encouraging them to ask questions or discuss topics they're interested in."
There is no other context and no set system prompt to ask for this.
Y'all getting this too? This same is on Qwen3-235B-A22B, no quants; full FP16
| 2025-04-29T02:09:40 |
SashaUsesReddit
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kad08f
| false | null |
t3_1kad08f
|
/r/LocalLLaMA/comments/1kad08f/qwen_3_wants_to_respond_in_chinese_even_when_not/
| false | false | 16 |
{'enabled': True, 'images': [{'id': 'mBPrWaZPpeV48cYv5fmZu-5KV-QaZszleWlCvJadiW4', 'resolutions': [{'height': 64, 'url': 'https://preview.redd.it/w1hxndcymoxe1.png?width=108&crop=smart&auto=webp&s=389de9788e33bde9b96c338ef270f9b5cc32221c', 'width': 108}, {'height': 129, 'url': 'https://preview.redd.it/w1hxndcymoxe1.png?width=216&crop=smart&auto=webp&s=57335cee3bf17f575ded563e96c87467c72d73bf', 'width': 216}, {'height': 191, 'url': 'https://preview.redd.it/w1hxndcymoxe1.png?width=320&crop=smart&auto=webp&s=e193c9b5ea0d75b5907163763feedd9f16ab2d42', 'width': 320}, {'height': 382, 'url': 'https://preview.redd.it/w1hxndcymoxe1.png?width=640&crop=smart&auto=webp&s=e5baf8cb2ff308908a702359cf79feebd34ee806', 'width': 640}, {'height': 573, 'url': 'https://preview.redd.it/w1hxndcymoxe1.png?width=960&crop=smart&auto=webp&s=6ca39db54d06b3474c79cd7dae2c40a73fd84004', 'width': 960}, {'height': 645, 'url': 'https://preview.redd.it/w1hxndcymoxe1.png?width=1080&crop=smart&auto=webp&s=52ba88168e56b73cddd21119fe2a562e21d4552e', 'width': 1080}], 'source': {'height': 1047, 'url': 'https://preview.redd.it/w1hxndcymoxe1.png?auto=webp&s=9f2efbe748889e16c047a64885ce58bf5d08bf2a', 'width': 1752}, 'variants': {}}]}
|
||
Someone please make this
| 1 |
So after every new model drop, I find myself browsing reddit and twitter in order to gauge sentiment for any new model drop. I think it's really important to gauge the community's reaction when it comes to model performance - outside of just checking benchmarks.
If someone put together something that automatically scrapes the sentiment from certain twitter accounts (maybe 50-100) + certain reddit communities and processes + displays the consensus in some form, that would be amazing. I feel like lots of people would value this.
| 2025-04-29T02:20:06 |
https://www.reddit.com/r/LocalLLaMA/comments/1kad7cc/someone_please_make_this/
|
cobalt1137
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kad7cc
| false | null |
t3_1kad7cc
|
/r/LocalLLaMA/comments/1kad7cc/someone_please_make_this/
| false | false |
self
| 1 | null |
OpenArc 1.0.3: Vision has arrrived, plus Qwen3!
| 1 |
[removed]
| 2025-04-29T02:27:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1kadc98/openarc_103_vision_has_arrrived_plus_qwen3/
|
Echo9Zulu-
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kadc98
| false | null |
t3_1kadc98
|
/r/LocalLLaMA/comments/1kadc98/openarc_103_vision_has_arrrived_plus_qwen3/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'dY_MHdD96fpFxFEVQJ8UvVyC_Yqm-C0hyS7sTCMZ5Mc', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/EEVrvoujAk9w4OeWoqMYygdotdKhdZz5qDMePd4a2KA.jpg?width=108&crop=smart&auto=webp&s=8ab05db3c982e9fec4b7520aaca06a8152dda3da', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/EEVrvoujAk9w4OeWoqMYygdotdKhdZz5qDMePd4a2KA.jpg?width=216&crop=smart&auto=webp&s=761e4a21c07a968ccf1b9aa4bc78760cd19c4ea7', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/EEVrvoujAk9w4OeWoqMYygdotdKhdZz5qDMePd4a2KA.jpg?width=320&crop=smart&auto=webp&s=44d340062796cc45c83ff9601a5fa88ca49f0af7', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/EEVrvoujAk9w4OeWoqMYygdotdKhdZz5qDMePd4a2KA.jpg?width=640&crop=smart&auto=webp&s=511cdf8fc8059700f4ecd54ed6c2ceaf84e44bd9', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/EEVrvoujAk9w4OeWoqMYygdotdKhdZz5qDMePd4a2KA.jpg?width=960&crop=smart&auto=webp&s=f4cfae1248104f26d07414069b922adde2edb6d2', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/EEVrvoujAk9w4OeWoqMYygdotdKhdZz5qDMePd4a2KA.jpg?width=1080&crop=smart&auto=webp&s=5d1f5838b0f8201d3ed74f57d3ad4d42e5fecd6a', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/EEVrvoujAk9w4OeWoqMYygdotdKhdZz5qDMePd4a2KA.jpg?auto=webp&s=fdc63790ee1b5e36d78ee90c677d3922e82d1fac', 'width': 1200}, 'variants': {}}]}
|
Qwen 3 (4B to 14B) the model that's sorry but dumb
| 0 |
And the bad joke starts again. Another "super launch", with very high Benchmark scores. In practice: terrible model in multilingualism; spends hundreds of tokens (in "thinking" mode) to answer trivial things. And the most shocking thing: if you don't "think" you get confused and answer wrong.
I've never seen a community more (...) to fall for hype. I include myself in this, I'm a muggle. Anyway, thanks Qwen, for Llama4.2.
| 2025-04-29T02:47:05 |
https://www.reddit.com/r/LocalLLaMA/comments/1kadpp1/qwen_3_4b_to_14b_the_model_thats_sorry_but_dumb/
|
sunomonodekani
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kadpp1
| false | null |
t3_1kadpp1
|
/r/LocalLLaMA/comments/1kadpp1/qwen_3_4b_to_14b_the_model_thats_sorry_but_dumb/
| false | false |
self
| 0 | null |
Which is smarter: Qwen 3 14B, or Qwen 3 30B A3B?
| 52 |
I'm running with 16GB of VRAM, and I was wondering which of these two models are smarter.
| 2025-04-29T02:56:58 |
https://www.reddit.com/r/LocalLLaMA/comments/1kadwbs/which_is_smarter_qwen_3_14b_or_qwen_3_30b_a3b/
|
RandumbRedditor1000
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kadwbs
| false | null |
t3_1kadwbs
|
/r/LocalLLaMA/comments/1kadwbs/which_is_smarter_qwen_3_14b_or_qwen_3_30b_a3b/
| false | false |
self
| 52 | null |
Asked tiny Qwen3 to make a self portrait using Matplotlib:
| 36 | 2025-04-29T03:03:17 |
https://www.reddit.com/gallery/1kae0j8
|
JLeonsarmiento
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kae0j8
| false | null |
t3_1kae0j8
|
/r/LocalLLaMA/comments/1kae0j8/asked_tiny_qwen3_to_make_a_self_portrait_using/
| false | false | 36 | null |
||
If I tell any Qwen3 model on oLlama to "Write me an extremely long essay about dogs", it goes into an infinite loop when it tries to finish the essay.
| 2 |
Per title. It's usually a "Note" section at the end, sometimes includes "Final Word Count", sometimes a special statement about dogs, but it just keeps looping spitting out a few minor variations of a short section of similar text forever. Once , the 4b version broke out of this and just started printing lines of only ''' forever.
What gives? Is there something wrong with how oLlama is setting these models up?
| 2025-04-29T03:19:15 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaeat3/if_i_tell_any_qwen3_model_on_ollama_to_write_me/
|
Mooseral
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaeat3
| false | null |
t3_1kaeat3
|
/r/LocalLLaMA/comments/1kaeat3/if_i_tell_any_qwen3_model_on_ollama_to_write_me/
| false | false |
self
| 2 | null |
Qwen3 AWQ Support Confirmed (PR Check)
| 22 |
[https://github.com/casper-hansen/AutoAWQ/pull/751](https://github.com/casper-hansen/AutoAWQ/pull/751)
Confirmed Qwen3 support added. Nice.
| 2025-04-29T03:35:13 |
https://www.reddit.com/r/LocalLLaMA/comments/1kael9w/qwen3_awq_support_confirmed_pr_check/
|
Acceptable-State-271
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kael9w
| false | null |
t3_1kael9w
|
/r/LocalLLaMA/comments/1kael9w/qwen3_awq_support_confirmed_pr_check/
| false | false |
self
| 22 |
{'enabled': False, 'images': [{'id': 'b_ptr20W6CfZajV2y9zq8EnErkaFh6hhyCyD0ndrlcI', 'resolutions': [{'height': 54, 'url': 'https://external-preview.redd.it/mXqmID4ePWVa0xpe8rDdeW-VCnFOkFtTyKBfAcGhNwE.jpg?width=108&crop=smart&auto=webp&s=e289166910610fb3c548060555249b3824cdd4c4', 'width': 108}, {'height': 108, 'url': 'https://external-preview.redd.it/mXqmID4ePWVa0xpe8rDdeW-VCnFOkFtTyKBfAcGhNwE.jpg?width=216&crop=smart&auto=webp&s=db40c9f9b4bb18e48c2120a081c0d78fdd41d380', 'width': 216}, {'height': 160, 'url': 'https://external-preview.redd.it/mXqmID4ePWVa0xpe8rDdeW-VCnFOkFtTyKBfAcGhNwE.jpg?width=320&crop=smart&auto=webp&s=f15a75ec73e98d58e991dfff5de9da20ddf5bfed', 'width': 320}, {'height': 320, 'url': 'https://external-preview.redd.it/mXqmID4ePWVa0xpe8rDdeW-VCnFOkFtTyKBfAcGhNwE.jpg?width=640&crop=smart&auto=webp&s=0c5846bcd28a41825ab67782566d3eaff5e4e1dd', 'width': 640}, {'height': 480, 'url': 'https://external-preview.redd.it/mXqmID4ePWVa0xpe8rDdeW-VCnFOkFtTyKBfAcGhNwE.jpg?width=960&crop=smart&auto=webp&s=819fa646401c352317802dbe35b21d65c01d0196', 'width': 960}, {'height': 540, 'url': 'https://external-preview.redd.it/mXqmID4ePWVa0xpe8rDdeW-VCnFOkFtTyKBfAcGhNwE.jpg?width=1080&crop=smart&auto=webp&s=ce1b603cf9092b710e2f9ac46e103ab532f1ced2', 'width': 1080}], 'source': {'height': 600, 'url': 'https://external-preview.redd.it/mXqmID4ePWVa0xpe8rDdeW-VCnFOkFtTyKBfAcGhNwE.jpg?auto=webp&s=f8fd404a553f35f5438cc2d6968880b0a631d778', 'width': 1200}, 'variants': {}}]}
|
Qwen 235B A22B vs Sonnet 3.7 Thinking - Pokémon UI
| 32 | 2025-04-29T03:40:10 |
sirjoaco
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaeobg
| false | null |
t3_1kaeobg
|
/r/LocalLLaMA/comments/1kaeobg/qwen_235b_a22b_vs_sonnet_37_thinking_pokémon_ui/
| false | false | 32 |
{'enabled': True, 'images': [{'id': 'TQrPi6LefsKkTf1bSaTv3h_JG9Tj-Azssg1Lcs3JmIA', 'resolutions': [{'height': 67, 'url': 'https://preview.redd.it/0uiwnqfy2pxe1.jpeg?width=108&crop=smart&auto=webp&s=f6d9f60575640a567be32bea5891ffd5f8921d54', 'width': 108}, {'height': 134, 'url': 'https://preview.redd.it/0uiwnqfy2pxe1.jpeg?width=216&crop=smart&auto=webp&s=dc0eb31528e7dc6fa5737718ba3c5b6a1962e812', 'width': 216}, {'height': 199, 'url': 'https://preview.redd.it/0uiwnqfy2pxe1.jpeg?width=320&crop=smart&auto=webp&s=c73d397c27d2aca8c9da09320622f4b7198df6cf', 'width': 320}, {'height': 399, 'url': 'https://preview.redd.it/0uiwnqfy2pxe1.jpeg?width=640&crop=smart&auto=webp&s=9bcfcdbef30612e9527a48230fd9d870062b416e', 'width': 640}, {'height': 599, 'url': 'https://preview.redd.it/0uiwnqfy2pxe1.jpeg?width=960&crop=smart&auto=webp&s=4045e8aa7bbbe1e9d08a07349e0d857dd883100d', 'width': 960}, {'height': 674, 'url': 'https://preview.redd.it/0uiwnqfy2pxe1.jpeg?width=1080&crop=smart&auto=webp&s=238a197a3c7f238cc4bfdf4335e2cebef634705c', 'width': 1080}], 'source': {'height': 1888, 'url': 'https://preview.redd.it/0uiwnqfy2pxe1.jpeg?auto=webp&s=291d3ce2e58c9d4f428e039234f18bc0640a5a2f', 'width': 3024}, 'variants': {}}]}
|
|||
Can i run some LLM?
| 3 |
My pc have this video card:
Model: RTX 4060 Ti
Memory: 8 GB
CUDA: Activado (versión 12.8).
something for code with Aider? or maybe other general things? or is useless?
Also i have:
xxxxxxx@fedora:~$ free -h
total used free shared buff/cache available
Mem: 30Gi 4,0Gi 23Gi 90Mi 3,8Gi 26Gi
Thanks!
| 2025-04-29T03:41:27 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaep5b/can_i_run_some_llm/
|
9acca9
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaep5b
| false | null |
t3_1kaep5b
|
/r/LocalLLaMA/comments/1kaep5b/can_i_run_some_llm/
| false | false |
self
| 3 | null |
Completely turn off "Thinking" of Qwen3 with this ollama chat template (modelfile)
| 2 |
You can use this ollama modelfile with custom chat template which will completely disable the <think> block, like this:
https://preview.redd.it/87lkv3q94pxe1.png?width=1903&format=png&auto=webp&s=1e5782c1602e67f9366df921780eb946049d5a64
This edited chat template works by insert an empty think block before model generate any text, so you don't have to use any scripts to remove the think block, it just works like a normal non-reasoning model.
`ollama create Qwen3-4B-IQ4_XS_NoThink -f modelfile`
Here is the Modelfile:
FROM Qwen3-4B-IQ4_XS.gguf
TEMPLATE """{{- if .Messages }}
{{- if or .System .Tools }}<|im_start|>system
{{- if .System }}
{{ .System }}
{{- end }}
{{- if .Tools }}
# Tools
You may call one or more functions to assist with the user query.
You are provided with function signatures within <tools></tools> XML tags:
<tools>
{{- range .Tools }}
{"type": "function", "function": {{ .Function }}}
{{- end }}
</tools>
For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call>
{{- end }}<|im_end|>
{{ end }}
{{- range $i, $_ := .Messages }}
{{- $last := eq (len (slice $.Messages $i)) 1 -}}
{{- if eq .Role "user" }}<|im_start|>user
{{ .Content }}<|im_end|>
{{ else if eq .Role "assistant" }}<|im_start|>assistant
<think>
</think>
{{ if .Content }}{{ .Content }}
{{- else if .ToolCalls }}<tool_call>
{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
{{ end }}</tool_call>
{{- end }}{{ if not $last }}<|im_end|>
{{ end }}
{{- else if eq .Role "tool" }}<|im_start|>user
<tool_response>
{{ .Content }}
</tool_response><|im_end|>
{{ end }}
{{- if and (ne .Role "assistant") $last }}<|im_start|>assistant
<think>
</think>
{{ end }}
{{- end }}
{{- else }}
{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
<think>
</think>
{{ end }}{{ .Response }}{{ if .Response }}<|im_end|>{{ end }}"""
PARAMETER repeat_penalty 1
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.6
PARAMETER top_k 20
PARAMETER top_p 0.95
| 2025-04-29T03:53:34 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaewsh/completely_turn_off_thinking_of_qwen3_with_this/
|
AaronFeng47
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaewsh
| false | null |
t3_1kaewsh
|
/r/LocalLLaMA/comments/1kaewsh/completely_turn_off_thinking_of_qwen3_with_this/
| false | false | 2 | null |
|
Llama CPP - Poorer Speed with Qwen3-30b-3ba than I expected. Help?
| 1 |
[removed]
| 2025-04-29T03:59:53 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaf0p5/llama_cpp_poorer_speed_with_qwen330b3ba_than_i/
|
EmPips
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaf0p5
| false | null |
t3_1kaf0p5
|
/r/LocalLLaMA/comments/1kaf0p5/llama_cpp_poorer_speed_with_qwen330b3ba_than_i/
| false | false |
self
| 1 | null |
Request for assistance with Ollama issue
| 5 |
Hello all -
I downloaded Qwen3 14b, and 30b and was going through the motions of testing them for personal use when I ended up walking away for 30 mins. I came back, and ran the 14b model and ran into an issue that now replicates across all local models, including non-Qwen models which is an error stating "llama runner process has terminated: GGML\_ASSERT(tensor->op == GGML\_OP\_UNARY) failed".
Normally, I can run these models with no issues, and even the Qwen3 models were running quickly. Any ideas for a novice on where I should be looking to try to fix it?
| 2025-04-29T04:06:19 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaf4va/request_for_assistance_with_ollama_issue/
|
MusukoRising
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaf4va
| false | null |
t3_1kaf4va
|
/r/LocalLLaMA/comments/1kaf4va/request_for_assistance_with_ollama_issue/
| false | false |
self
| 5 | null |
1 gorilla vs 100 men benchmark
| 1 |
[removed]
| 2025-04-29T04:10:18 |
https://www.reddit.com/gallery/1kaf7ah
|
Healthy_Albatross_73
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaf7ah
| false | null |
t3_1kaf7ah
|
/r/LocalLLaMA/comments/1kaf7ah/1_gorilla_vs_100_men_benchmark/
| false | false | 1 | null |
|
Deepseek R2
| 1 |
[removed]
| 2025-04-29T04:16:37 |
https://www.reddit.com/r/LocalLLaMA/comments/1kafb2d/deepseek_r2/
|
TankProfessional8947
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kafb2d
| false | null |
t3_1kafb2d
|
/r/LocalLLaMA/comments/1kafb2d/deepseek_r2/
| false | false |
self
| 1 | null |
Browser-based AI with 3D avatars — MLC LLM + Xenova pipelines!
| 1 |
[removed]
| 2025-04-29T04:16:38 |
https://www.reddit.com/r/LocalLLaMA/comments/1kafb2r/browserbased_ai_with_3d_avatars_mlc_llm_xenova/
|
xelacleo
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kafb2r
| false | null |
t3_1kafb2r
|
/r/LocalLLaMA/comments/1kafb2r/browserbased_ai_with_3d_avatars_mlc_llm_xenova/
| false | false |
self
| 1 | null |
Qwen3 8B FP16 - asked for 93 items, got 93 items.
| 256 |
tried many times - alwas exact list length.
Without using minItems.
in my daily work this is a breakthrough!
| 2025-04-29T04:16:48 |
secopsml
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kafb6h
| false | null |
t3_1kafb6h
|
/r/LocalLLaMA/comments/1kafb6h/qwen3_8b_fp16_asked_for_93_items_got_93_items/
| false | false | 256 |
{'enabled': True, 'images': [{'id': 'Zhj1lUT1Ho4m6_w9Q699_VB94F81_y9L90WEwf9UYg0', 'resolutions': [{'height': 103, 'url': 'https://preview.redd.it/61wa8u139pxe1.png?width=108&crop=smart&auto=webp&s=d9492adca5d1f94a9021f8dfb6fb748dbc1861f7', 'width': 108}, {'height': 206, 'url': 'https://preview.redd.it/61wa8u139pxe1.png?width=216&crop=smart&auto=webp&s=f6c60265acf6dab722342169b8ece4a6cf7a12a5', 'width': 216}, {'height': 306, 'url': 'https://preview.redd.it/61wa8u139pxe1.png?width=320&crop=smart&auto=webp&s=a65461281e7f7b4a1dd6e8c658cfcade8429c50f', 'width': 320}, {'height': 612, 'url': 'https://preview.redd.it/61wa8u139pxe1.png?width=640&crop=smart&auto=webp&s=13d7490e62b1815f603b33c4e952eaaceb7f52aa', 'width': 640}, {'height': 918, 'url': 'https://preview.redd.it/61wa8u139pxe1.png?width=960&crop=smart&auto=webp&s=1dc5e0e5dccc8419049ebf7385d3b40380ecce29', 'width': 960}], 'source': {'height': 919, 'url': 'https://preview.redd.it/61wa8u139pxe1.png?auto=webp&s=c3c4a9a467095486d92bc48d52412567c54bcce0', 'width': 961}, 'variants': {}}]}
|
||
Slow Qwen3-30B-A3B speed on 4090, can't utilize gpu properly
| 8 |
I tried unsloth Q4 gguf with ollama and llama.cpp, both can't utilize my gpu properly, only running at 120 watts
I tought it's ggufs problem, then I downloaded Q4KM gguf from ollama library, same issue
Any one knows what may cause the issue? I tried turn on and off kv cache, zero difference
| 2025-04-29T04:34:48 |
https://www.reddit.com/r/LocalLLaMA/comments/1kafm0l/slow_qwen330ba3b_speed_on_4090_cant_utilize_gpu/
|
AaronFeng47
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kafm0l
| false | null |
t3_1kafm0l
|
/r/LocalLLaMA/comments/1kafm0l/slow_qwen330ba3b_speed_on_4090_cant_utilize_gpu/
| false | false |
self
| 8 | null |
Which is best among these 3 qwen models
| 10 | 2025-04-29T04:43:53 |
Namra_7
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kafrae
| false | null |
t3_1kafrae
|
/r/LocalLLaMA/comments/1kafrae/which_is_best_among_these_3_qwen_models/
| false | false | 10 |
{'enabled': True, 'images': [{'id': '_u_j6Eu9PbOEDIWoq9BYXHPh47JdsCBsT3_Gx5lrLCk', 'resolutions': [{'height': 174, 'url': 'https://preview.redd.it/jhvkprmiepxe1.png?width=108&crop=smart&auto=webp&s=96d95f262f69dc2354cb73d325f72ddb8811af35', 'width': 108}, {'height': 348, 'url': 'https://preview.redd.it/jhvkprmiepxe1.png?width=216&crop=smart&auto=webp&s=eedf9b7041db593d1e7defe5e04968e6f94193fe', 'width': 216}, {'height': 516, 'url': 'https://preview.redd.it/jhvkprmiepxe1.png?width=320&crop=smart&auto=webp&s=98f478eafd26e48df7714dfc9238b5c9c14827ec', 'width': 320}, {'height': 1032, 'url': 'https://preview.redd.it/jhvkprmiepxe1.png?width=640&crop=smart&auto=webp&s=2e2c72a0a951ec0584732fa0d2dee4e73f68e5f5', 'width': 640}, {'height': 1548, 'url': 'https://preview.redd.it/jhvkprmiepxe1.png?width=960&crop=smart&auto=webp&s=7df1e3fa1727597f1c1ce4e86405a63b21aa8599', 'width': 960}, {'height': 1742, 'url': 'https://preview.redd.it/jhvkprmiepxe1.png?width=1080&crop=smart&auto=webp&s=a40edf7c027c71c347ca2e352dda6dbdafa73f45', 'width': 1080}], 'source': {'height': 1742, 'url': 'https://preview.redd.it/jhvkprmiepxe1.png?auto=webp&s=25ce2d12b10ad7b4c9b1909d3cc9ad19b54d529b', 'width': 1080}, 'variants': {}}]}
|
|||
Is it possible to do FAST image generation on a laptop
| 5 |
I am exhibiting at a tradeshow soon and I thought a fun activation could be instant-printed trading cards with them as a super hero/pixar etc.
Is there any local image gen with decent results that can run on a laptop (happy to purchase a new laptop). It needs to be FAST though - max 10 seconds (even that is pushing it).
Love to hear if it's possible
| 2025-04-29T04:45:11 |
https://www.reddit.com/r/LocalLLaMA/comments/1kafs0q/is_it_possible_to_do_fast_image_generation_on_a/
|
Plane_Garbage
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kafs0q
| false | null |
t3_1kafs0q
|
/r/LocalLLaMA/comments/1kafs0q/is_it_possible_to_do_fast_image_generation_on_a/
| false | false |
self
| 5 | null |
Qwen3 0.6B on Android runs flawlessly
| 259 |
I recently released v0.8.6 for ChatterUI, just in time for the Qwen 3 drop:
https://github.com/Vali-98/ChatterUI/releases/latest
So far the models seem to run fine out of the gate, and generation speeds are very optimistic for 0.6B-4B, and this is by far the smartest small model I have used.
| 2025-04-29T04:52:48 |
https://v.redd.it/3snum824gpxe1
|
----Val----
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kafwa7
| false |
{'reddit_video': {'bitrate_kbps': 5000, 'dash_url': 'https://v.redd.it/3snum824gpxe1/DASHPlaylist.mpd?a=1748494382%2CNTRhMTY1NzkyYTMxMzVjYWRjNGZjNDZlOTYwMTg0NTVjZTk0ZjNkZTEyNjc4ZmQ5NjVkZTkyOWQ0YWE5ZDEwZg%3D%3D&v=1&f=sd', 'duration': 21, 'fallback_url': 'https://v.redd.it/3snum824gpxe1/DASH_1080.mp4?source=fallback', 'has_audio': True, 'height': 1920, 'hls_url': 'https://v.redd.it/3snum824gpxe1/HLSPlaylist.m3u8?a=1748494382%2CNDE1MGEzZmIyZTRhZTRiODg0M2Y3YjYxOTU4NjNkYmNhYWZkMTAzN2RkNTk4NTE4NmI5MjA3ZmM1YWQ1ODcxMw%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/3snum824gpxe1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 864}}
|
t3_1kafwa7
|
/r/LocalLLaMA/comments/1kafwa7/qwen3_06b_on_android_runs_flawlessly/
| false | false | 259 |
{'enabled': False, 'images': [{'id': 'bXNhcjB5dTNncHhlMW--9zt028XHo6Q9w18q7qbHfuIyDOGMqIn7YhcX7Vou', 'resolutions': [{'height': 216, 'url': 'https://external-preview.redd.it/bXNhcjB5dTNncHhlMW--9zt028XHo6Q9w18q7qbHfuIyDOGMqIn7YhcX7Vou.png?width=108&crop=smart&format=pjpg&auto=webp&s=73b4e3df0ca5bc2961f426b539c101104417224e', 'width': 108}, {'height': 432, 'url': 'https://external-preview.redd.it/bXNhcjB5dTNncHhlMW--9zt028XHo6Q9w18q7qbHfuIyDOGMqIn7YhcX7Vou.png?width=216&crop=smart&format=pjpg&auto=webp&s=9d8506a288e04f551f095330906b3b2ae1da2b06', 'width': 216}, {'height': 640, 'url': 'https://external-preview.redd.it/bXNhcjB5dTNncHhlMW--9zt028XHo6Q9w18q7qbHfuIyDOGMqIn7YhcX7Vou.png?width=320&crop=smart&format=pjpg&auto=webp&s=d7ca0c60592bda9758c3fdeae051d02f6f8cda36', 'width': 320}, {'height': 1280, 'url': 'https://external-preview.redd.it/bXNhcjB5dTNncHhlMW--9zt028XHo6Q9w18q7qbHfuIyDOGMqIn7YhcX7Vou.png?width=640&crop=smart&format=pjpg&auto=webp&s=b7fcf7c5421725ceea54a9157860aa7eec25c870', 'width': 640}], 'source': {'height': 1756, 'url': 'https://external-preview.redd.it/bXNhcjB5dTNncHhlMW--9zt028XHo6Q9w18q7qbHfuIyDOGMqIn7YhcX7Vou.png?format=pjpg&auto=webp&s=5a4e2dd410bd96773cbc9031232cd62f4b4c3858', 'width': 790}, 'variants': {}}]}
|
|
Qwen3-30B-A3B runs at 12-15 tokens-per-second on CPU
| 903 |
CPU: AMD Ryzen 9 7950x3d
RAM: 32 GB
I am using the UnSloth Q6\_K version of Qwen3-30B-A3B ([Qwen3-30B-A3B-Q6\_K.gguf · unsloth/Qwen3-30B-A3B-GGUF at main](https://huggingface.co/unsloth/Qwen3-30B-A3B-GGUF/blob/main/Qwen3-30B-A3B-Q6_K.gguf))
| 2025-04-29T05:07:07 |
https://v.redd.it/k27mtpenipxe1
|
AlgorithmicKing
|
v.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kag4er
| false |
{'reddit_video': {'bitrate_kbps': 2400, 'dash_url': 'https://v.redd.it/k27mtpenipxe1/DASHPlaylist.mpd?a=1748495241%2CMWQ3OTMwMGUyMDE5NjFkYWUwYzlkNTdjYjQ3MzNmN2UwOGU4NzE3Njc2M2MyN2Q2ZmU5YzAzNTg2YjIzNDc0OQ%3D%3D&v=1&f=sd', 'duration': 109, 'fallback_url': 'https://v.redd.it/k27mtpenipxe1/DASH_720.mp4?source=fallback', 'has_audio': True, 'height': 720, 'hls_url': 'https://v.redd.it/k27mtpenipxe1/HLSPlaylist.m3u8?a=1748495241%2CMTM1YTQ0NWRkMDVhMzdlMTQ5YjlkYmNiMDRlNTI0ZDQzNWMxMGQyNGM5NjhjZjVhMjdlNGE3ZjM0YmFiMTE2MQ%3D%3D&v=1&f=sd', 'is_gif': False, 'scrubber_media_url': 'https://v.redd.it/k27mtpenipxe1/DASH_96.mp4', 'transcoding_status': 'completed', 'width': 1160}}
|
t3_1kag4er
|
/r/LocalLLaMA/comments/1kag4er/qwen330ba3b_runs_at_1215_tokenspersecond_on_cpu/
| false | false | 903 |
{'enabled': False, 'images': [{'id': 'M2wycXVrZ25pcHhlMaaUXUGYAurpwjopJY1zZmzeHwQa0qRuSYTlAz0bqREA', 'resolutions': [{'height': 67, 'url': 'https://external-preview.redd.it/M2wycXVrZ25pcHhlMaaUXUGYAurpwjopJY1zZmzeHwQa0qRuSYTlAz0bqREA.png?width=108&crop=smart&format=pjpg&auto=webp&s=7cb8a8719364c6a56a3bf46731048621b012e3f3', 'width': 108}, {'height': 134, 'url': 'https://external-preview.redd.it/M2wycXVrZ25pcHhlMaaUXUGYAurpwjopJY1zZmzeHwQa0qRuSYTlAz0bqREA.png?width=216&crop=smart&format=pjpg&auto=webp&s=c0e49e6db62517c4e309c6ef72a9f3c04a7f032d', 'width': 216}, {'height': 198, 'url': 'https://external-preview.redd.it/M2wycXVrZ25pcHhlMaaUXUGYAurpwjopJY1zZmzeHwQa0qRuSYTlAz0bqREA.png?width=320&crop=smart&format=pjpg&auto=webp&s=239aeacff56475947bc439d3dd3d69b1ab7b5978', 'width': 320}, {'height': 397, 'url': 'https://external-preview.redd.it/M2wycXVrZ25pcHhlMaaUXUGYAurpwjopJY1zZmzeHwQa0qRuSYTlAz0bqREA.png?width=640&crop=smart&format=pjpg&auto=webp&s=40babe56f7926eb4c9dbb8e3ee5546bc5c3d6063', 'width': 640}, {'height': 595, 'url': 'https://external-preview.redd.it/M2wycXVrZ25pcHhlMaaUXUGYAurpwjopJY1zZmzeHwQa0qRuSYTlAz0bqREA.png?width=960&crop=smart&format=pjpg&auto=webp&s=8572a3447b768cce5baee94871982a88b3148172', 'width': 960}, {'height': 670, 'url': 'https://external-preview.redd.it/M2wycXVrZ25pcHhlMaaUXUGYAurpwjopJY1zZmzeHwQa0qRuSYTlAz0bqREA.png?width=1080&crop=smart&format=pjpg&auto=webp&s=b2524cbdf9e6015ea53739bd8ce0e1b683d4d8aa', 'width': 1080}], 'source': {'height': 984, 'url': 'https://external-preview.redd.it/M2wycXVrZ25pcHhlMaaUXUGYAurpwjopJY1zZmzeHwQa0qRuSYTlAz0bqREA.png?format=pjpg&auto=webp&s=3540da72b44e6df7b30c567d3101554ef99886d2', 'width': 1586}, 'variants': {}}]}
|
|
abstract-agent: Locally hosted AI Agent Python Tool To Generate Novel Research Hypothesis + Titles + Abstracts
| 4 |
Base foundation v0.1 - Testing with the new qwen 600m model and getting pretty decent outputs
## Features
- Multi-agent pipeline: breakdown, critique, synthesize, innovate, and polish
- Pulls from public sources: arXiv, Semantic Scholar, EuropePMC, Crossref, DOAJ, bioRxiv, medRxiv, OpenAlex, PubMed
- Scores, ranks, and summarizes literature
- Uses Ollama for summarization and novelty checks
- Final output is a clean, human-readable panel with stats / insights
| 2025-04-29T05:44:45 |
https://github.com/tegridydev/abstract-agent
|
tegridyblues
|
github.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kagov3
| false | null |
t3_1kagov3
|
/r/LocalLLaMA/comments/1kagov3/abstractagent_locally_hosted_ai_agent_python_tool/
| false | false |
default
| 4 | null |
Fine tuning rune Qwen 3 0.6b
| 8 |
Has anyone tried to find tune Qwen 3 0.6b? I am seeing you guys running it everyone, I wonder if I could run a fine tuned version as well.
Thanks
| 2025-04-29T05:48:49 |
https://www.reddit.com/r/LocalLLaMA/comments/1kagr26/fine_tuning_rune_qwen_3_06b/
|
Effective_Head_5020
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kagr26
| false | null |
t3_1kagr26
|
/r/LocalLLaMA/comments/1kagr26/fine_tuning_rune_qwen_3_06b/
| false | false |
self
| 8 | null |
Qwen Team - Huge Thank You! Can you also release Qwen2.5 Max?!
| 1 |
[removed]
| 2025-04-29T05:52:47 |
https://www.reddit.com/r/LocalLLaMA/comments/1kagt7o/qwen_team_huge_thank_you_can_you_also_release/
|
nite2k
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kagt7o
| false | null |
t3_1kagt7o
|
/r/LocalLLaMA/comments/1kagt7o/qwen_team_huge_thank_you_can_you_also_release/
| false | false |
self
| 1 | null |
Amount of parameters vs Quantization
| 1 |
Which is more important for pure conversation? no mega intelligence that has a doctorate in neruo sciences needed, just plain pure fun coversation.
| 2025-04-29T05:57:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1kagvll/amount_of_parameters_vs_quantization/
|
XPEZNAZ
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kagvll
| false | null |
t3_1kagvll
|
/r/LocalLLaMA/comments/1kagvll/amount_of_parameters_vs_quantization/
| false | false |
self
| 1 | null |
Inquiry about Unsloth's quantization methods
| 4 |
I noticed that Unsloth has added a UD version in GGUF quantization. I would like to ask, under the same size, is the UD version better? For example, is the quality of UD-Q3\_K\_XL.gguf higher than Q4\_KM and IQ4\_XS?
| 2025-04-29T05:58:44 |
https://www.reddit.com/r/LocalLLaMA/comments/1kagwas/inquiry_about_unsloths_quantization_methods/
|
Few_Professional6859
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kagwas
| false | null |
t3_1kagwas
|
/r/LocalLLaMA/comments/1kagwas/inquiry_about_unsloths_quantization_methods/
| false | false |
self
| 4 | null |
Abliterated Qwen3 when?
| 5 |
I know it's a bit too soon but god its fast.
And please make the 30b a3b first.
| 2025-04-29T06:03:18 |
https://www.reddit.com/r/LocalLLaMA/comments/1kagyta/abliterated_qwen3_when/
|
AlgorithmicKing
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kagyta
| false | null |
t3_1kagyta
|
/r/LocalLLaMA/comments/1kagyta/abliterated_qwen3_when/
| false | false |
self
| 5 | null |
Running Qwen 3 on Zimacube pro and RTX pro 6000
| 3 |
Maybe at this point the question is cliché
But it would be great to get SOTA llm at full power running locally for an affordable price
There's a new NAS called Zimacube pro, it looks like a new personal cloud with server options, they have a lot of capabilities and it looks great
But what about installing the new RTX pro 6000 on that zimacube pro?
Is it there a boilerplate of requirements for SOTA models? (Deepseek r1 671B, ot this new Qwen3)
Assuming you won't have bottleneck,what you guys think about using Zimacube pro with 2 RTX pro 6000 for server, cloud, multimedia services and unlimited llm in your home?
I really want to learn about that, so I would appreciate your thoughts
| 2025-04-29T06:11:22 |
Xoloshibu
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kah33d
| false | null |
t3_1kah33d
|
/r/LocalLLaMA/comments/1kah33d/running_qwen_3_on_zimacube_pro_and_rtx_pro_6000/
| false | false | 3 |
{'enabled': True, 'images': [{'id': 'zUE_NJw1iYFrCVcwDW8G79gpoZudHXZvNOep8cFUF1w', 'resolutions': [{'height': 96, 'url': 'https://preview.redd.it/p1872ep4upxe1.jpeg?width=108&crop=smart&auto=webp&s=0c37a3e8a0547e73c468a7737f01e9a712888990', 'width': 108}, {'height': 192, 'url': 'https://preview.redd.it/p1872ep4upxe1.jpeg?width=216&crop=smart&auto=webp&s=0ce0a2cbb630874169725c3e8d87e5657b6fd9b5', 'width': 216}, {'height': 285, 'url': 'https://preview.redd.it/p1872ep4upxe1.jpeg?width=320&crop=smart&auto=webp&s=1161896f64a03aa5cb7e8238e85e2881a675252a', 'width': 320}, {'height': 571, 'url': 'https://preview.redd.it/p1872ep4upxe1.jpeg?width=640&crop=smart&auto=webp&s=f361d4e7e152dc8b0c3157bbf5551b334ab1099f', 'width': 640}, {'height': 856, 'url': 'https://preview.redd.it/p1872ep4upxe1.jpeg?width=960&crop=smart&auto=webp&s=cbe79a13679550dfc4a35f1ae3ec94b211cb99d4', 'width': 960}, {'height': 963, 'url': 'https://preview.redd.it/p1872ep4upxe1.jpeg?width=1080&crop=smart&auto=webp&s=d69351ce685ecd19fb6345e0bd6bc8742394df9c', 'width': 1080}], 'source': {'height': 1146, 'url': 'https://preview.redd.it/p1872ep4upxe1.jpeg?auto=webp&s=2269a3a1180cb47292f1d6eca237e6e87b32ed19', 'width': 1284}, 'variants': {}}]}
|
||
Qwen3 Plays DnD while Chatgpt DMs.
| 4 |
https://m.twitch.tv/cm0rduck
Trying Qwen3 32b with dnd, with chatgpt DMing
| 2025-04-29T06:12:44 |
https://www.reddit.com/r/LocalLLaMA/comments/1kah3sz/qwen3_plays_dnd_while_chatgpt_dms/
|
Spare-Ad-4810
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kah3sz
| false | null |
t3_1kah3sz
|
/r/LocalLLaMA/comments/1kah3sz/qwen3_plays_dnd_while_chatgpt_dms/
| false | false |
self
| 4 |
{'enabled': False, 'images': [{'id': '-nis-KMHsOnz-UjoGNvC_5M866Ui_bkxycNuXqxqtNE', 'resolutions': [{'height': 108, 'url': 'https://external-preview.redd.it/ayveQl_Bx3B7Ex1HKYmUNJEHTn4ld0Yuo_aLKh-UBAc.jpg?width=108&crop=smart&auto=webp&s=d366b3ba03903ea422e927620c56117d38fb6344', 'width': 108}, {'height': 216, 'url': 'https://external-preview.redd.it/ayveQl_Bx3B7Ex1HKYmUNJEHTn4ld0Yuo_aLKh-UBAc.jpg?width=216&crop=smart&auto=webp&s=3cadae6d78f577c40a69f730db5f85f7325feb9f', 'width': 216}], 'source': {'height': 300, 'url': 'https://external-preview.redd.it/ayveQl_Bx3B7Ex1HKYmUNJEHTn4ld0Yuo_aLKh-UBAc.jpg?auto=webp&s=3f7367bc6e666611cded6723961581f360455576', 'width': 300}, 'variants': {}}]}
|
How are applications like Base44 built?
| 2 |
Hi all,
In short, I’m asking about applications that create other applications from a prompt — how does the layer work that translates the prompt into the API that builds the app?
From what I understand, after the prompt is processed, it figures out which components need to be built: GUI, backend, third-party APIs, etc.
So, in short, how is this technically built?
| 2025-04-29T06:19:15 |
https://www.reddit.com/r/LocalLLaMA/comments/1kah77a/how_are_applications_like_base44_built/
|
umen
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kah77a
| false | null |
t3_1kah77a
|
/r/LocalLLaMA/comments/1kah77a/how_are_applications_like_base44_built/
| false | false |
self
| 2 | null |
Need help with creating a dataset for fine-tuning embeddings model
| 4 |
So I've come across dozens of posts where they've fine tuned embeddings model for getting a better contextual embedding for a particular subject.
So I've been trying to do something and I'm not sure how to create a pair label / contrastive learning dataset.
From many videos i saw they've taken a base model and they've extracted the embeddings and calculate cosine and use a threshold to assign labels but thisbmethod won't it bias the model to the base model lowkey sounds like distillation ot a model .
Second one was to use some rule based approach and key words to find out the similarity but the dataset is in a crass format to find the keywords.
Third is to use a LLM to label using prompting and some knowledge to find out the relation and label it.
I've ran out of ideas and people who have done this before pls tell ur ideas and guide me on how to do.
| 2025-04-29T06:36:56 |
https://www.reddit.com/r/LocalLLaMA/comments/1kahgjh/need_help_with_creating_a_dataset_for_finetuning/
|
Sanjuej
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kahgjh
| false | null |
t3_1kahgjh
|
/r/LocalLLaMA/comments/1kahgjh/need_help_with_creating_a_dataset_for_finetuning/
| false | false |
self
| 4 | null |
Qwen 3 - The "thinking" is very slow.
| 0 |
Anyone else experiencing this? Is displaying the "thinking" super slow. Like the system is just running slow or something. Been happening all day.
Any suggestions? Sign out and then back in?
| 2025-04-29T06:41:47 |
https://www.reddit.com/r/LocalLLaMA/comments/1kahj1j/qwen_3_the_thinking_is_very_slow/
|
Glittering-Cancel-25
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kahj1j
| false | null |
t3_1kahj1j
|
/r/LocalLLaMA/comments/1kahj1j/qwen_3_the_thinking_is_very_slow/
| false | false |
self
| 0 | null |
DAY-1 Qwen3-235B (Q4, CPU-only) vs DeepSeek R1-671B (Hybrid Q2, GPU+CPU) — first-look speed test, NOT a fair fight… yet!
| 1 |
[removed]
| 2025-04-29T06:43:23 |
https://www.reddit.com/r/LocalLLaMA/comments/1kahjtu/day1_qwen3235b_q4_cpuonly_vs_deepseek_r1671b/
|
texasdude11
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kahjtu
| false | null |
t3_1kahjtu
|
/r/LocalLLaMA/comments/1kahjtu/day1_qwen3235b_q4_cpuonly_vs_deepseek_r1671b/
| false | false |
self
| 1 |
{'enabled': False, 'images': [{'id': 'QhJbBPSnRtznzTOWUxSMCU9luTckmqssEdLnM87rEH0', 'resolutions': [{'height': 81, 'url': 'https://external-preview.redd.it/x4LzdXglFFnUgOeJkhC1oXKrW3mfsSMVaaHAjt7jX-k.jpg?width=108&crop=smart&auto=webp&s=3779ae4f43969f3049346ee7d56396d3e264d41f', 'width': 108}, {'height': 162, 'url': 'https://external-preview.redd.it/x4LzdXglFFnUgOeJkhC1oXKrW3mfsSMVaaHAjt7jX-k.jpg?width=216&crop=smart&auto=webp&s=aed1eab3a104ee56c75ef543e6568eabab8495e3', 'width': 216}, {'height': 240, 'url': 'https://external-preview.redd.it/x4LzdXglFFnUgOeJkhC1oXKrW3mfsSMVaaHAjt7jX-k.jpg?width=320&crop=smart&auto=webp&s=f12f06bdaedd51eca593878699a20f3cc85241ee', 'width': 320}], 'source': {'height': 360, 'url': 'https://external-preview.redd.it/x4LzdXglFFnUgOeJkhC1oXKrW3mfsSMVaaHAjt7jX-k.jpg?auto=webp&s=5bc70f1ba247cbe75a7be945333e23c6fbfea419', 'width': 480}, 'variants': {}}]}
|
I am VERY impressed by qwen3 4B (q8q4 gguf version)
| 59 |
I usually test models reasoning using a few "not in any dataset" logic problems.
Up until the thinking models came along, only "huge" models could solve "some" of those problems in one shot.
Today I wanted to see how a heavily quantized (q8q4) small model as Qwen3 4B performed.
To my surprise, it gave the right answer and even the thinking was linear and very good.
You can find my quants here: [https://huggingface.co/ZeroWw/Qwen3-4B-GGUF](https://huggingface.co/ZeroWw/Qwen3-4B-GGUF)
| 2025-04-29T07:32:14 |
https://www.reddit.com/r/LocalLLaMA/comments/1kai88d/i_am_very_impressed_by_qwen3_4b_q8q4_gguf_version/
|
Robert__Sinclair
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kai88d
| false | null |
t3_1kai88d
|
/r/LocalLLaMA/comments/1kai88d/i_am_very_impressed_by_qwen3_4b_q8q4_gguf_version/
| false | false |
self
| 59 |
{'enabled': False, 'images': [{'id': 'cTwZak_-WLVO582OdmnEuAJO-IOY9oSAiwK2Gywi34Y', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/tZ2emz_Vby2-LHI51Af1yqWnuqQcJKx0JFm79_vRJAA.jpg?width=108&crop=smart&auto=webp&s=b57c3897b9f7befa75e9155c7743b7a049c61d4c', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/tZ2emz_Vby2-LHI51Af1yqWnuqQcJKx0JFm79_vRJAA.jpg?width=216&crop=smart&auto=webp&s=c2ce85d7b09e0dbd8aa906c9bd54faa6538c8cc2', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/tZ2emz_Vby2-LHI51Af1yqWnuqQcJKx0JFm79_vRJAA.jpg?width=320&crop=smart&auto=webp&s=6205bcfeed4aba233071174887115b70248ad478', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/tZ2emz_Vby2-LHI51Af1yqWnuqQcJKx0JFm79_vRJAA.jpg?width=640&crop=smart&auto=webp&s=e12c6186cd9cc218bef148026fe7351ea489dc57', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/tZ2emz_Vby2-LHI51Af1yqWnuqQcJKx0JFm79_vRJAA.jpg?width=960&crop=smart&auto=webp&s=16fc8ac995245416960c6ec8b0a253e4d535cb39', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/tZ2emz_Vby2-LHI51Af1yqWnuqQcJKx0JFm79_vRJAA.jpg?width=1080&crop=smart&auto=webp&s=df265a79c46606b92a3642193900c7ce6222fc75', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/tZ2emz_Vby2-LHI51Af1yqWnuqQcJKx0JFm79_vRJAA.jpg?auto=webp&s=2f58d961a23b241f48e3355c2b563f960acaf71a', 'width': 1200}, 'variants': {}}]}
|
How can i download qwen android app
| 1 |
[removed]
| 2025-04-29T07:47:47 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaifu0/how_can_i_download_qwen_android_app/
|
FRENLYFROK
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaifu0
| false | null |
t3_1kaifu0
|
/r/LocalLLaMA/comments/1kaifu0/how_can_i_download_qwen_android_app/
| false | false |
self
| 1 | null |
M4 max chip for AI development
| 1 |
[removed]
| 2025-04-29T07:48:01 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaifz1/m4_max_chip_for_ai_development/
|
Similar_Tangerine142
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaifz1
| false | null |
t3_1kaifz1
|
/r/LocalLLaMA/comments/1kaifz1/m4_max_chip_for_ai_development/
| false | false |
self
| 1 | null |
Psychological and opinion manipulation by LLM-powered bots was unleashed on CMV during a University of Zurich study.
| 1 | 2025-04-29T07:57:34 |
https://www.reddit.com/r/changemyview/comments/1k8b2hj/meta_unauthorized_experiment_on_cmv_involving/
|
Sea_Sympathy_495
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaikjn
| false | null |
t3_1kaikjn
|
/r/LocalLLaMA/comments/1kaikjn/psychological_and_opinion_manipulation_by/
| false | false |
default
| 1 | null |
|
Qwen3 after the hype
| 281 |
Now that I hope the initial hype has subsided, how are each models really?
* [](https://huggingface.co/Qwen/Qwen3-235B-A22B)
* [](https://huggingface.co/Qwen/Qwen3-30B-A3B)
* [](https://huggingface.co/Qwen/Qwen3-32B)
* [](https://huggingface.co/Qwen/Qwen3-14B)
* [](https://huggingface.co/Qwen/Qwen3-8B)
* [](https://huggingface.co/Qwen/Qwen3-4B)
* [](https://huggingface.co/Qwen/Qwen3-1.7B)
* [Qwen/Qwen3-0.6B ](https://huggingface.co/Qwen/Qwen3-0.6B)
Beyond the benchmarks, how are they really feeling according to you in terms of coding, creative, brainstorming and thinking? What are the strengths and weaknesses?
| 2025-04-29T08:05:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaioin/qwen3_after_the_hype/
|
Cheap_Concert168no
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaioin
| false | null |
t3_1kaioin
|
/r/LocalLLaMA/comments/1kaioin/qwen3_after_the_hype/
| false | false |
self
| 281 |
{'enabled': False, 'images': [{'id': '2CPXSIzkp22xYPsTVpgsp4OcDlEliyzHHGoKpPeBFBs', 'resolutions': [{'height': 58, 'url': 'https://external-preview.redd.it/9XgXXrBl6uCwW0eE-n9K4sfzB5YuRIuJJ-W66FyTl_w.jpg?width=108&crop=smart&auto=webp&s=89957ddc3e0ceb4136c276d3d85968c69109c147', 'width': 108}, {'height': 116, 'url': 'https://external-preview.redd.it/9XgXXrBl6uCwW0eE-n9K4sfzB5YuRIuJJ-W66FyTl_w.jpg?width=216&crop=smart&auto=webp&s=e4a44b08f3f4a0dced4c58658a931f1249f694f2', 'width': 216}, {'height': 172, 'url': 'https://external-preview.redd.it/9XgXXrBl6uCwW0eE-n9K4sfzB5YuRIuJJ-W66FyTl_w.jpg?width=320&crop=smart&auto=webp&s=a3de8cd502780f769e1d3da532275ec4fd60c53f', 'width': 320}, {'height': 345, 'url': 'https://external-preview.redd.it/9XgXXrBl6uCwW0eE-n9K4sfzB5YuRIuJJ-W66FyTl_w.jpg?width=640&crop=smart&auto=webp&s=4f7cd146bad9c79b3890a9abf11391b109bd9776', 'width': 640}, {'height': 518, 'url': 'https://external-preview.redd.it/9XgXXrBl6uCwW0eE-n9K4sfzB5YuRIuJJ-W66FyTl_w.jpg?width=960&crop=smart&auto=webp&s=602260ad00758e767180485deb7d9ee48f77343e', 'width': 960}, {'height': 583, 'url': 'https://external-preview.redd.it/9XgXXrBl6uCwW0eE-n9K4sfzB5YuRIuJJ-W66FyTl_w.jpg?width=1080&crop=smart&auto=webp&s=de3b7dea2a2861846122e38b8c470914a9487010', 'width': 1080}], 'source': {'height': 648, 'url': 'https://external-preview.redd.it/9XgXXrBl6uCwW0eE-n9K4sfzB5YuRIuJJ-W66FyTl_w.jpg?auto=webp&s=cc4bd6610df3c9a4f902269a1812e41e7888b950', 'width': 1200}, 'variants': {}}]}
|
This DVD can code better than me
| 1 |
[removed]
| 2025-04-29T08:08:43 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaiq2x/this_dvd_can_code_better_than_me/
|
tralalala2137
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaiq2x
| false | null |
t3_1kaiq2x
|
/r/LocalLLaMA/comments/1kaiq2x/this_dvd_can_code_better_than_me/
| false | false | 1 | null |
|
Tried running Qwen3-32B and Qwen3-30B-A3B on my Mac M2 Ultra. The 3B-active MoE doesn’t feel as fast as I expected.
| 4 | 2025-04-29T08:09:13 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaiqbf/tried_running_qwen332b_and_qwen330ba3b_on_my_mac/
|
Known-Classroom2655
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaiqbf
| false | null |
t3_1kaiqbf
|
/r/LocalLLaMA/comments/1kaiqbf/tried_running_qwen332b_and_qwen330ba3b_on_my_mac/
| false | false | 4 | null |
||
Qwen3 30b a3b q4_K_M performance on M1 Ultra
| 1 |
Through Ollama, on M1 Ultra 128GB RAM I got following values:
`response_token/s: 29.95`
`prompt_token/s: 362.26`
`total_duration: 72708617792`
`load_duration: 12474000`
`prompt_eval_count: 1365`
`prompt_tokens: 1365`
`prompt_eval_duration: 3768006375`
`eval_count: 2064`
`completion_tokens: 2064`
`eval_duration: 68912612667`
`approximate_total: "0h1m12s"`
`total_tokens: 3429`
Not what I expected (I thought its gonna run faster). For reference, I rerun the query with gemma model and got something along response\_token/s \~65 and prompt\_token/s: \~1600 (similar prompt\_tokens and eval\_count, so its not caused by thinking and degradation).
So, even though its a3b, its more than 2x slower for generation than gemma 4b model, and its more than 4x slower for prompt processing than gemma 4b. Is it normal?
| 2025-04-29T08:11:58 |
https://www.reddit.com/r/LocalLLaMA/comments/1kairl6/qwen3_30b_a3b_q4_k_m_performance_on_m1_ultra/
|
One_Key_8127
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kairl6
| false | null |
t3_1kairl6
|
/r/LocalLLaMA/comments/1kairl6/qwen3_30b_a3b_q4_k_m_performance_on_m1_ultra/
| false | false |
self
| 1 | null |
Oh damn even qwen 3 now it's censored for Taiwan question
| 1 | 2025-04-29T08:12:02 |
Independent-Wind4462
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kairm5
| false | null |
t3_1kairm5
|
/r/LocalLLaMA/comments/1kairm5/oh_damn_even_qwen_3_now_its_censored_for_taiwan/
| false | false | 1 |
{'enabled': True, 'images': [{'id': 'fB1ZUeTetUJlC3hzNe07edJEgotDohVqEA1szMYShN8', 'resolutions': [{'height': 134, 'url': 'https://preview.redd.it/s0hw0uinfqxe1.jpeg?width=108&crop=smart&auto=webp&s=1b164031c92b44bf2a8ef93b6be0a2010c152527', 'width': 108}, {'height': 269, 'url': 'https://preview.redd.it/s0hw0uinfqxe1.jpeg?width=216&crop=smart&auto=webp&s=619e8adbe8d94360909a3246660bc91445378883', 'width': 216}, {'height': 399, 'url': 'https://preview.redd.it/s0hw0uinfqxe1.jpeg?width=320&crop=smart&auto=webp&s=e0f0b315f8a60a1cc35732d74c8b913ed44ddf32', 'width': 320}, {'height': 799, 'url': 'https://preview.redd.it/s0hw0uinfqxe1.jpeg?width=640&crop=smart&auto=webp&s=3b92306b4ba2ba8d37b185e35e6799dde4812f0f', 'width': 640}, {'height': 1199, 'url': 'https://preview.redd.it/s0hw0uinfqxe1.jpeg?width=960&crop=smart&auto=webp&s=de5dbbc39172f06e56c8538b9ddef198a7d6838a', 'width': 960}], 'source': {'height': 1344, 'url': 'https://preview.redd.it/s0hw0uinfqxe1.jpeg?auto=webp&s=386e454308cf415fd45c4663ad14ca676e3087e6', 'width': 1076}, 'variants': {}}]}
|
|||
Qwen 3
| 0 | 2025-04-29T08:14:07 |
United_Dimension_46
|
i.redd.it
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaiskv
| false | null |
t3_1kaiskv
|
/r/LocalLLaMA/comments/1kaiskv/qwen_3/
| false | false | 0 |
{'enabled': True, 'images': [{'id': 'bIqc5ySMW1k17pD935UTLVHL1uhh5912lCk6CcNyThY', 'resolutions': [{'height': 192, 'url': 'https://preview.redd.it/ejusb831gqxe1.jpeg?width=108&crop=smart&auto=webp&s=2e2b9f3dea9c524c870ac23e7e9bd4552c8af874', 'width': 108}, {'height': 384, 'url': 'https://preview.redd.it/ejusb831gqxe1.jpeg?width=216&crop=smart&auto=webp&s=34ea59bc339f9824c4b9f26b33dddc2d07e758c1', 'width': 216}, {'height': 568, 'url': 'https://preview.redd.it/ejusb831gqxe1.jpeg?width=320&crop=smart&auto=webp&s=7132c19941a76589fa6492febf62f21fb4a9d6eb', 'width': 320}, {'height': 1137, 'url': 'https://preview.redd.it/ejusb831gqxe1.jpeg?width=640&crop=smart&auto=webp&s=093471aaa1e817032220f2e86478ff183b6f2060', 'width': 640}, {'height': 1706, 'url': 'https://preview.redd.it/ejusb831gqxe1.jpeg?width=960&crop=smart&auto=webp&s=d5d5c8f6480f83e02df5c2f773c3d8e943f23395', 'width': 960}, {'height': 1920, 'url': 'https://preview.redd.it/ejusb831gqxe1.jpeg?width=1080&crop=smart&auto=webp&s=68fe4c15d72e006c45cd2e3643be8ecc70e88720', 'width': 1080}], 'source': {'height': 3840, 'url': 'https://preview.redd.it/ejusb831gqxe1.jpeg?auto=webp&s=96a31ca70f3531c4835a4de289c88bd28d785313', 'width': 2160}, 'variants': {}}]}
|
|||
No Qwen 3 on lmarena?
| 4 |
Do you remember how it was with 2.5 and QwQ? Did they add it later after the release?
| 2025-04-29T08:23:03 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaiwqs/no_qwen_3_on_lmarena/
|
jacek2023
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaiwqs
| false | null |
t3_1kaiwqs
|
/r/LocalLLaMA/comments/1kaiwqs/no_qwen_3_on_lmarena/
| false | false |
self
| 4 | null |
【Evaluation】Best practice for evaluating Qwen3 !!
| 1 |
[removed]
| 2025-04-29T08:27:28 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaiypo/evaluationbest_practice_for_evaluating_qwen3/
|
Flashy_Sign2764
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaiypo
| false | null |
t3_1kaiypo
|
/r/LocalLLaMA/comments/1kaiypo/evaluationbest_practice_for_evaluating_qwen3/
| false | false | 1 | null |
|
is second state legit ? can get to run models on lm studio
| 3 | 2025-04-29T08:44:17 |
https://www.reddit.com/gallery/1kaj6o9
|
agx3x2
|
reddit.com
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaj6o9
| false | null |
t3_1kaj6o9
|
/r/LocalLLaMA/comments/1kaj6o9/is_second_state_legit_can_get_to_run_models_on_lm/
| false | false | 3 | null |
||
The QWEN 3 score does not match the actual experience
| 63 |
qwen 3 is great, but is it a bit of an exaggeration? Is QWEN3-30B-A3B really stronger than Deepseek v660 0324B? I've found that deepseek has a better ability to work in any environment, for example in cline\roo code\SillyTavern, deepseek can do it with ease, but qwen3-30b-a3b can't, even the more powerful qwen3-235b-a22b can't, it usually gets lost in context, don't you think? What are your use cases?
| 2025-04-29T08:50:58 |
https://www.reddit.com/r/LocalLLaMA/comments/1kaj9v7/the_qwen_3_score_does_not_match_the_actual/
|
vihv
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kaj9v7
| false | null |
t3_1kaj9v7
|
/r/LocalLLaMA/comments/1kaj9v7/the_qwen_3_score_does_not_match_the_actual/
| false | false |
self
| 63 | null |
Qwen 3 0.6b first impressions
| 1 |
[removed]
| 2025-04-29T08:53:58 |
https://www.reddit.com/r/LocalLLaMA/comments/1kajb9j/qwen_3_06b_first_impressions/
|
ThundeR_StaR
|
self.LocalLLaMA
| 1970-01-01T00:00:00 | 0 |
{}
|
1kajb9j
| false | null |
t3_1kajb9j
|
/r/LocalLLaMA/comments/1kajb9j/qwen_3_06b_first_impressions/
| false | false |
self
| 1 | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.