Alibaba unveils Qwen3-Max, its first trillion-parameter model

Alibaba unveiled Qwen3-Max at its annual Apsara Conference, describing it as the company’s largest language model to date, with more than a trillion parameters. The model was released in separate Instruct and Thinking modes, following the pattern Alibaba had already established across the smaller Qwen3 line of pairing a fast non-reasoning mode with a slower mode that works through extended chains of reasoning before answering.

Alibaba reported a score of 69.6 on SWE-bench for the Instruct version, which it said put the model on par with some leading closed-source rivals, alongside strong results on Tau2-Bench, an evaluation of conversational agents handling multi-turn tool use. Unlike most of the Qwen family, which Alibaba has released with open weights and credited with hundreds of millions of downloads and hundreds of thousands of derivative fine-tunes, Qwen3-Max was made available only through Alibaba Cloud’s API, without published weights.

The closed-weight choice marked a departure for a company that had built much of its reputation, and a meaningful share of the open-source ecosystem’s dependence on Chinese labs, on releasing frontier-class models openly. Alibaba did not detail its reasoning for withholding weights on its largest model while continuing to open-source smaller ones, but the split reflected a broader pattern among Chinese labs of using open releases to build developer mindshare and adoption while reserving their most capable, most expensive-to-train models for commercial API revenue. The announcement came a week into a run of significant Chinese model releases that month, alongside updates from DeepSeek and Tencent.