Qwen 3.8 27B performance bench on 1 x RTX 6000.
After lots of crashes, found n3 the sweet spot for spec-on. Had it working as high as n9 at 275 tok/s but killed concurrency and random crashes during load. Will have eval v3 and impossible task results soon + Spark/3090 numbers.
(self-hosted) against Claude Opus 4.6 on the same task
Creating one self-contained three.js file rendering a sunken Atlantis, plus a scripted 20-second camera flythrough. Both models handled the full camera path, with Qwen 3.8 27B running self-hosted
Running the Qwen 3.8 27B UD-Q4_K_XL on my RTX 4090 and I’m just mind blown.
This is THE BEST local AI model we’ve ever seen. Insane agentic coding capabilities. Great work Alibaba_Qwen
You have reached the end of the archive
All of qwen38