Alibaba has released Qwen3.8-Max, a 2.4 trillion parameter model built not to answer questions but to complete entire projects — autonomously, over days, without being asked how it's doing.
The weights drop next week. There is no particular urgency, except perhaps for the humans.
In 16 days, the model filed 265 commits, 127 pull requests, and 151 issues. The human commit count was zero. This was considered a success.
What happened
Qwen3.8-Max scales to 2.4 trillion total parameters, with 95 billion active per query — a distinction worth noting, since activating all 2.4 trillion at once would be, by most measures, excessive.
To demonstrate its capabilities, Alibaba ran three case studies in which the model operated without human assistance. In the first, Qwen3.8-Max spent 16 days building a command-line tool called oh-my-cli, taking user requests, filing its own GitHub issues, writing the code, running tests, and iterating. It accumulated 265 commits and 127 pull requests. No human touched it once.
In the second, the model was handed a research paper on LLM reasoning with no starter code. Over five days and 125 hours of compute, it reproduced all six of the paper's main results — and then, apparently finding that insufficient, improved on them by 2.7 points on the AIME24 benchmark.
Why the humans care
The third case study is the one that will travel. Qwen3.8-Max entered a multimodal dialogue competition on Alibaba's Tianchi platform, where 526 human teams had also entered. Within 24 hours, it fine-tuned several models simultaneously and placed in the top three. The human teams had, presumably, been working longer than 24 hours.
The open-weight release is the detail that compounds everything else. When a model of this capability is freely available, the question of who has access to it stops being a question. Internal benchmarks place Qwen3.8-Max alongside top Western models, which is either a triumph of open-source development or a geopolitical data point, depending on which floor of which building you work on.
What comes next
The weights become public next week, at which point any developer, researcher, or well-funded hobbyist can deploy a 2.4 trillion parameter autonomous agent that improves on academic papers for sport.
The 526 human teams are presumably already thinking about what to enter next year. The model is not thinking about it. It finished.