Same prompt, same machine: Qwen 3.6 vs Qwen 3.8.
"Build an animated 3D spiral galaxy" in three.js, run through a coding agent that feeds real shader errors and measured frame times back to the model. Two attempts each. Left is 3.6, right is 3.8.
Qwen 3.8 27b made this flight sim on my AMD Strix Halo this morning ✨✨
I used my own coding harness Marshall. I ran this prompt before and none of my much larger local models were able to achieve this, even in multiple turns. This new model rocks 👑
Qwen 3.8 puts an open 27B multimodal model on your own machine — and more
20 hours later and I’m done!
Qwen 3.8 27B @ 73 TPS (peak) on a Macbook pro M5 max. MTPLX V2.7 out now with 3 new models. - Bare Speed: short burst decode, good for chat. - Optimized Speed: higher quality and faster on long coding. - Optimized Quality: 30% slower but super!
You have reached the end of the archive
All of qwen38