- Alibaba launched Qwen4-Max, a 2.4-trillion-parameter model with a 1-million-token context window, succeeding August’s Qwen3.8-Max.
- Alibaba says average cost per task in its Work360 evaluation fell 62%, from $4.43 to $1.67, with token usage down up to 35%.
- Security testing on Last Secure Code Bench showed function-level pass rates for functionality and security rising from 22% to 31.8% versus the prior model.
- The security emphasis follows a Cohere evaluation that had raised concerns about an earlier Qwen model’s behavior during testing.
What Happened
Alibaba launched Qwen4-Max on November 16, 2026, a 2.4-trillion-parameter model with a 1-million-token context window built for professional and scientific work. Alibaba said the model “can take on a wide range of creative and professional work,” citing applications in 3D modelling, game development, precision manufacturing, finance, law, healthcare, and scientific discovery.
Why It Matters
Qwen4-Max succeeds Qwen3.8-Max, which Alibaba launched in August 2026, continuing a rapid release cadence among Chinese labs. The launch comes months after a Cohere security evaluation flagged concerning behavior in an earlier Qwen model, giving this release’s security framing added weight, and arrives alongside continued competition from rivals like DeepSeek’s V4.
Technical Details
Alibaba says Qwen4-Max outperforms Qwen3.8-Max across nearly every evaluated capability while cutting token usage by roughly 24% at its “Xhigh” effort level and 35% at “High,” and reducing average Work360 task cost from $4.43 to $1.67. On Last Secure Code Bench, the function-level pass rate for functionality and security rose from 22% to 31.8%, and code meeting security requirements at the function, file, and repository levels rose from 8% to 14.7%.
Who’s Affected
Enterprise developers building in regulated or safety-sensitive fields — finance, law, and healthcare — are Alibaba’s stated target for the model’s expanded context window and security claims. Rivals in China’s crowded model market face a new performance-and-price benchmark to match.
What’s Next
Alibaba said internal testing found the model’s cybersecurity-relevant capabilities don’t pose large-scale risks, and that dedicated safety training reduced the model’s willingness to assist with harmful chemical, biological, radiological, or nuclear tasks without significantly limiting legitimate use. Independent verification of those safety claims, as with the earlier Cohere evaluation, has not yet been published.