Six months ago, Claude Opus 4.6 sat near the top of the AI world.
Now a 27B open model is knocking on the same door. Qwen 3.8-27B is: → Free to download. → Open under Apache 2.0. → Capable of running locally. → Multimodal across text, images, and video. → Packing
Grok 4.6 + Seedance 2.5 or GPT-5.6 + Seedance 2.5 or Gemini 3.7 + Seedance 2.5...
Can be used for this. For a lower price, we can use Qwen 3.8 Max, Kimi K3, and DeepSeek V4. For quality, we can use Opus 5, and Claude Fable 5.
Qwen 3.8 27B running locally on a single RTX 5090 is getting ridiculous.
It beat Opus 4.8 in my benchmarks and hit ~200 tok/s in some tests.
This was made with a ROCmFP3 version of Qwen 3.8 27B on Strix Halo.
All made by talking to the Strix Framework Desktop remotely while I am at dinner. Now working on mobile touch controls.
You have reached the end of the archive
All of qwen38