Okay, I am fairly confident in my hypothesis now.
The secret sauce behind Qwen 3.8 27B becomes almost immediately evident during testing. It is not the training data. In fact, I doubt any SFT was involved at all. The model was simply allowed GRPO with a more liberal reasoning
Qwen 3.8 27B in full BF16 on one 96GB M3 Ultra Mac Studio.
54.74GB weights • 58.09GB peak 21.51 tok/s with native MTP speculative decoding—63% faster than no drafter in my test. The video uses real timestamped stream chunks. Exact Hugging Face recipe ↓
Qwen 3.8 is a game-changer!
It's performing on par with the latest Opus model. Seriously impressed with the capabilities here.
Finally tested Alibaba_Qwen Qwen 3.8 27B on two RTX 5090 GPUs.
🔥 Getting around 120–130 tokens/sec with vLLM, UnslothAI NVFP4, native MTP, and a 150K context target. The coding output looks very good.
You have reached the end of the archive
All of qwen38