Qwen just dropped Qwen3.8-27B and the numbers are wild.
If you've been running 3.5 or 3.6 locally, this is the first "must-upgrade" jump in a while — especially if you care about agentic coding, tool use, or computer-use agents. Here's the headline story in raw benchmark points
Alibaba's Qwen 3.8 Max ran a coding project alone for 16 days straight.
265 commits. 127 pull requests. It took in feedback, dispatched tasks, tested, and self-corrected. No human prompting each step. That's the new bar for AI agents. → Beat Fable 5 and GPT 5.6 on OS World,
One shotted (one line prompt) this game on my laptop locally using Qwen 3.8 27B IQ XXS
Run Qwen 3.8 27B (dense) 14 tokens/sec on FREE 2x Nvidia T4 GPUs in under 5 mins!
While everyone is crying about running out of VRAM on their RTX 4090s/3090s, I just ran Qwen 3.8 27B (with MTP and 120k context) on a completely free cloud cluster using silicon from 2018. Kaggle
You have reached the end of the archive
All of qwen38