By Hamza Faisal
Alibaba has launched Qwen4-Max, its most powerful artificial intelligence model yet, claiming it beats leading Western rivals on coding tests — while keeping the flagship's weights closed, a sharp break from the open approach that made the Qwen family famous.
The launch, first reported by Bloomberg on Monday, makes Qwen4-Max the top model in Alibaba's new Qwen4 series. It is built on a sparse mixture-of-experts design with a 2.4-trillion-parameter backbone, handles text, images and video, and offers a 1-million-token context window — enough to hold roughly an entire codebase in a single conversation.
The benchmark claims — and the fine print
Alibaba says Qwen4-Max scored 75.7 on DeepSWE v1.1, a test of real-world software engineering, putting it ahead of Anthropic's Claude Opus 5.5 and competitive with Moonshot AI's Kimi K3.1 Pro. Bloomberg's report frames the model as outperforming Anthropic's Claude Fable 5.1 on coding benchmarks.
The numbers come with the usual caveat: they are Alibaba's own results. Independent rankings tell a more mixed story, placing Qwen4-Max behind Kimi on some leaderboards even while highlighting its lower cost per task. On one live-session leaderboard, Qwen4-Max trailed Claude Fable 5.2 by a wide margin on task success — a reminder that vendor benchmarks are a starting point, not a verdict.
Where the model clearly wins is price. Alibaba set Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens — a fraction of Claude Fable 5.1's $10 and $50 — and says it uses about 50 per cent fewer tokens than its predecessor, Qwen3.8-Max, with more than 60 per cent lower average cost per task in internal tests.
Closed weights, open smaller models
The biggest surprise is what Alibaba did not release. Unlike earlier Qwen flagships, Qwen4-Max will not be available as downloadable weights — developers can only reach it through Alibaba's QwenCloud API, which is compatible with protocols from OpenAI and Anthropic.
That is a notable shift for the world's most downloaded open-weight model family. Alibaba says smaller models in the Qwen4 series will still be released as open weights, including a Qwen4-27B model expected to run on machines with 16GB of memory. The flagship itself, however, stays closed.
A quiet preview of Qwen4-Max had been running in Alibaba Cloud's Model Studio since November 11, available only through a Token Plan. That preview runs until November 25.
A crowded race in China
The launch lands in a busy few weeks for Chinese AI labs. Tencent has shipped Hunyuan 4, Moonshot published the open weights for its 2.8-trillion-parameter Kimi K3.1 Pro earlier this month, and DeepSeek continues to iterate on its own flagship line. All three are now competing directly on coding and reasoning while tightening control over their best models.
The closed-weights choice also sits awkwardly next to Alibaba chairman Joe Tsai's recent backing of open-source AI at China's Wuzhen Summit — a tension the company will have to explain as it asks developers to trust a flagship they cannot inspect or run themselves.
For now, Qwen4-Max is live on QwenCloud. Whether developers adopt it will depend less on Alibaba's benchmark charts than on how it performs on their own code.



