Qwen3.8-Max: A New Bar for Coding and Cowork
Points and comments are a snapshot, not live.
Qwen releases 2.4-trillion-parameter model with open weights coming next week.
Qwen 3.8-Max has 2.4 trillion parameters (95B active) and will be open-sourced next week, the first Max-class model to do so. The blog highlights three autonomous coding feats: building a self-evolving harness over 16 days (265 commits), reproducing and improving a research paper (+2.71 points on AIME24), and beating 87% of human teams in a 24-hour contest. For work tasks, it claims to match or exceed several harnesses and demonstrates end-to-end quant strategy development and chip design.
What commenters are saying
Several commenters note the model was already available as a preview on July 19. Simon Willison tests it for SVG generation and reports it forgot to include wheels in a bicycle drawing, though the reasoning trace mentioned them. Debate arises over whether AI will reduce work hours or simply increase workload; one commenter says they are busier with AI than before. Another notes Qwen3.6-27B is widely used locally, and hopes the new 3.8-27B improves on it.