China is leading the way in open-weight model distribution, with many of its top AI labs releasing their models online for anyone to download and run, provided they have the required hardware. The latest model to get this treatment is Alibaba’s Qwen3.8-Max, with the firm revealing that it plans to publish the weights on ModelScope and Hugging Face next week.
Alibaba’s shares jumped in response, with its Hong Kong-listed stock closing 7% up on Monday, followed by a 4.2% rise in pre-market trading in the United States.
A spokesperson for the firm highlighted that this is "the first time we will open-source the weights of a Qwen-Max-class model." This is the company’s most powerful model; it has previously released the weights for its lesser-powered budget models, while guarding access to the flagship product.
The trend among Chinese developers, however, is increasingly shifting to open-weight access. Beijing-based Moonshot AI also published the weights for its own 2.8-trillion-parameter model Kimi K3 on July 27.
While U.S. frontier AI labs keep their model weights strictly proprietary, the trend in China is to offer comparable results - albeit not as powerful in many regards - in a cheaper and freely downloadable package.
Qwen3.8-Max showcases capabilities on mid-to-long-horizon tasks
Alibaba’s Qwen3.8-Max features 2.4 trillion parameters and can process up to a million tokens per query. The model is being marketed as a powerful tool for jobs with longer time horizons, rather than chatbot-style interactions.
To showcase these capabilities, the developer assigned complex machine-learning and coding tasks. These included a multi-week software development project that resulted in 265 commits, 127 pull requests and 151 issues after roughly 16 days without any human intervention. Another test asked the model to cut the size of a cryptographic circuit chip; Qwen successfully reduced the physical size of the component by 81% by streamlining its number of logic gates from 8,298 to 678.
The model also has use cases in e-commerce, as displayed by a final test in which it was given ¥100,000 and tasked with earning as much as possible across a year-long simulation. The model managed to quadruple its capital in that time, finishing 152% higher than the previous iteration of Alibaba’s flagship model.
These tests were run in-house at Alibaba according to its own rules and procedures, and have not been independently audited.
How Qwen3.8-Max compares to the competition
Qwen’s capabilities are perhaps less impressive when pitted against frontier U.S. models. Alibaba’s own comparison table reveals that it outperforms in some task types, but lags behind in others. On Terminal Bench 2.1 - which tests models’ ability to complete complex, multi-part tasks - the model scored 86.6, placing it ahead of the flagship models from Anthropic, but still slightly behind OpenAI’s GPT-5.6 Sol at 88.8.
However, with a score of 93.0, Qwen beats all contenders on the PaperBench test, which looks at the ability to replicate machine learning research papers. Other benchmark results were less flattering: Alibaba’s model scored just 67.7 on coding test SWE-bench Pro, far behind Claude Fable 5 at 80.0.
When limited to domestic competitors, however, Qwen3.8-Max’s position improves. Bloomberg reported that it ranks above the much-lauded Kimi K3 - another open-weight model of comparable size - on several key benchmarks, although Moonshot’s model performs marginally better overall.
This nonetheless places Qwen3.8-Max among the top contenders for open-weight models, especially where tasks with longer time horizons are concerned.