What's this about?
Alibaba released Qwen3.8-Max, the most capable Qwen model to date: 2.4 trillion parameters with 95B active, available now via API, with the weights being open-sourced next week — the first time a Qwen-Max-class model goes open-weights. Instead of leading with benchmarks, the release leads with three autonomy stories: it built and ran a self-evolving CLI project for 16 days on its own (265 commits, 127 PRs), reproduced a research paper's results from scratch and then improved on them (+2.7 on AIME24), and beat 526 human teams in a live data-science contest within 24 hours.
What exactly does "2.4T parameters (95B active)" mean?
Parameters are the model's learned knobs — crudely, its brain size. This model is a "mixture of experts": of the 2.4 trillion total parameters, only about 95 billion fire for any given token, like a hospital with thousands of specialists where each patient only sees the relevant few. You get huge total capacity while paying (in compute) closer to a much smaller model. "Open weights" means anyone can download and run the trained model — a big deal at this scale, since Max-class Qwen models were previously API-only.
What's the community saying?
- simonw spotted a naming muddle: an earlier Qwen tweet described "Qwen3.8" differently, so what exactly ships open next week (Max vs the 27B) confused even close watchers.
- Excitement centers on the smaller Qwen3.8-27B also announced for open release — its predecessor is widely considered the best local model, so an upgrade matters more to most people than the 2.4T flagship.
- Skeptics note "self-evolves through feedback loops" sounds like distilled Claude Code behavior, and asked whether the autonomous-coding receipts survive scrutiny.
- Meta-cynicism: "once OpenAI and Anthropic are public, every such announcement becomes a reliable sell signal."