Uncensored version of Qwen3.8-27B on 5090

I wanted to run it on NVFP4 with 200K context. The prefill seems fast and it seems easy to use. There were no ready-made ones, so I baked it myself. ▫️ 2 types in NVFP4 version All 4 bits, The MLP is 4 bits and the attention and output layers are 8 bits.

rS_ローカルLLM×投資Aug 189

Qwen 3.8 max just beat GPT-5.6 sol max at actually using a computer.

And the benchmark isn't even the most interesting part. The numbers: → 2.4 trillion parameter Mixture-of-Experts model → Only ~95B parameters activate per token → 1M-token context window across text,

Julian Goldie SEOAug 18

Designing with Qwen 3.8 27B in I am in a bit of a shock.

It's a big deal. These one-shot designs are marginally different from Opus, there are still nuances to be fair, but it's dangerously close. Once cerebras runs this model north of 1000 tps 👉we

Tom KrchaAug 1824

Qwen 3.8 27b rebuilt the water physics with a video analysis skill.

It's still going, too. This is another intermediate result. Unreal!

Trent Zock-RobbinsAug 181