Reminder that local ai is also important.
This is Qwen 3.8 flash-next. The entire generation ran locally on my 5090 from the video to the actual code. I can’t wait for Qwen 4 series.
If you’re stuck on one llm engine assuming it’s the best, check again
Tested qwen 3.8 27b on 3090 (350w, 64k ctx) with 4 serving engines hyperqwen: fastest at everything except loading. 180 tok/s avg, and 448 tok/s total at 8 users tensorfold: same exl3 files as exllamav3 but
Like GPT-6 Astra, Qwen 3.8 27B can interpret model weights
Treat the mechanism as a merge gate: prove it in CI before the next demo reel.
A study showed that AI raised the math score of a small model from 7.9% to 48.1% by directly modifying the weights without learning.
But Banban and Jaehyun came out in one day. If you look at the conditions, the story is a little different. How much was really 'just looking at the weights' and how much was Python and model execution? It's organized in numbers 👇 The original text is MIT CSAIL
You have reached the end of the archive
All of qwen38