Official version five days after the preview
China's Alibaba Group on Nov. 16 officially released "Qwen4-Max," the new flagship of its "Qwen" AI models. Qwen's official X account announced it on the morning of Nov. 16 Japan time. It adopts a new "mixture of experts (MoE)" architecture that runs only part of the model for each query, has "native vision" to understand images directly, and can read 1 million tokens at once. The API became available on the company's cloud service "QwenCloud," and Alibaba said the model can also be tried on "Qwen Studio," "QwenWork" and the AI coding tool "Qoder."
๐ข Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching โ all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! ๐ - Blog: https://qwen.ai/blog?id=qwen4-max - API: https://www.qwencloud.com/models/qwen4-max - Token Plan: https://www.qwencloud.com/pricing/token-plan More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. ๐
On Nov. 11, Alibaba released a preview of Qwen4-Max only to subscribers of "Token Plan," a flat-rate plan for developers, and did not provide performance figures at the time (ORIENS's earlier report). The official version has the model ID "qwen4-max," and QwenCloud's documentation treats the preview "qwen4-max-preview" as a separate model. The documentation describes Qwen4-Max as the first model in the Qwen4 family, combining reasoning, image understanding and the execution of long-running tasks.
Prices cut from previous generation, cost per task down over 60% in internal evaluation
According to QwenCloud's model page, API pricing is $1.5 per million input tokens and $4.5 per million output tokens. "Implicit caching," which lowers costs when the same parts of a prompt, such as its opening, are reused, costs $0.075 per million tokens. Alibaba said all three prices are lower than those of the previous-generation Qwen3.8-Max, and that it cut implicit caching by 70%. According to QwenCloud's documentation, the model is offered in Singapore, Beijing and "Global" regions such as Frankfurt and Tokyo, with Beijing and Global sharing the same prices and Singapore having separate prices. OpenRouter, which offers models from multiple companies through a single API, also lists the model under the identifier qwen/qwen4-max at $1.50 for input and $4.50 for output (per million tokens).
On performance, Alibaba said that in its internal evaluation Qwen4-Max outperformed Qwen3.8-Max while using up to about 50% fewer tokens. It said the average cost to complete a single task fell by more than 60%. Both figures come from the company's own evaluation, and the post did not describe the tasks used or compare the model with other companies' models.
Alibaba says it beat the previous generation in two Arena categories
On the morning of the same day, Alibaba also posted results from the model evaluation site Arena. In "Code Arena: WebDev," which compares the ability to build web pages, it said the model placed sixth, up 74 points from Qwen3.8-Max. In "Agent Arena," an evaluation of agents in which AI carries out tasks on its own, it said the model outperformed Qwen3.8-Max by a net 7.5%, putting Alibaba fifth in the ranking by developer (lab). Both results were reported by Alibaba on its own X account.
1 million-token context, maximum output differs across documents
According to QwenCloud's documentation, the context length is 1 million tokens, matching the model page. The model page says that in addition to generating text, the model supports image and video understanding, and can use web search, tool calling and output in a fixed format. "Thinking," in which the model reasons before answering, is enabled by default, and the reasoning depth can be selected from "low," "high" and "xhigh," with high as the default. It supports APIs in the format of OpenAI's Chat Completions and Responses and Anthropic's Messages. However, it does not support additional training (fine-tuning) or deployment of the model in a user's own environment.
The amount the model can output at once is not consistent between the documentation and the model page. The documentation gives a maximum of 131,072 tokens without thinking and 262,144 tokens for thinking and the answer combined. The model page shows the maximum output as "393K."
Weights of smaller models to be released later
In the announcement post, Alibaba said it will release the weights of smaller Qwen4 models. It did not say whether it will release the weights of Qwen4-Max itself. The official developer account "Qwen Developers" also cited the new architecture, native vision and the 1 million-token context, and posted that more Qwen4 models are on the way.





