On 16 November 2026, Alibaba released Qwen4-Max, the most capable model in the Qwen family so far. The multimodal model handles text, images and video, offers a context window of one million tokens and is aimed mainly at coding, office work and agentic tasks. Alibaba puts the emphasis on efficiency: according to its own tests, Qwen4-Max needs significantly fewer tokens than its predecessor. Through the API, it costs 1.50 US dollars per million input tokens and 4.50 US dollars per million output tokens in Singapore.
What Alibaba announced about Qwen4-Max
According to the Alibaba Cloud press release, Qwen4-Max is a mixture-of-experts (MoE) model with 2.4 trillion total parameters. Only 71 billion of them are active per token. Alibaba says that is 25 percent fewer active parameters than Qwen3.8-Max, which is also a 2.4-trillion-parameter model. The new Qwen4 architecture reworks the MoE design across attention, residual connections and embeddings, among other areas.
The model has native vision capabilities, and Alibaba says it is particularly strong at coding, agentic tasks and visual reasoning. As an external reference point, the company cites sixth place on Code Arena, a public leaderboard for real-world coding tasks. It also points to strong results on Agents’ Last Exam, Vals Index, JobBench and Chartography, without giving specific scores for those tests in the release.
Efficiency: fewer tokens per task
The key claim is about cost. On Work360, Alibaba’s proprietary benchmark of demanding tasks across 721 specialised fields, Qwen4-Max outperforms Qwen3.8-Max overall, according to the company. At the “High” reasoning effort, it uses 35 percent fewer tokens, and Alibaba says the average cost per task drops by around 60 percent.
These figures should be read with care. Work360 is Alibaba’s own test, and no independent evaluation is available yet. For developers, the direction still matters: with reasoning models, thinking tokens often make up a large part of the bill. A model that writes fewer tokens for the same task can be much cheaper even at similar per-token prices. Whether that holds for your own use case can only be answered with your own tests.
Pricing, regions and API access
The Alibaba Cloud Model Studio documentation lists the model ID qwen4-max. According to the docs, prices are the same with and without “thinking”:
- Singapore (International): 1.50 US dollars per million input tokens and 4.50 US dollars per million output tokens, plus a free quota of one million tokens valid for 90 days.
- Frankfurt, Virginia, Tokyo, Hong Kong (Global) and Beijing: 1.413 US dollars for input and 4.24 US dollars for output per million tokens.
- Cache: according to the press release, input with an implicit cache hit costs 0.075 US dollars per million tokens.
The context window is 1,000,000 tokens. Without thinking, the model can produce up to 131,072 output tokens, and reasoning plus answer can reach 262,144 tokens combined. Thinking is on by default, and the default reasoning effort is “high”, with “low” and “xhigh” as alternatives.
Useful for existing projects: the model is available through the OpenAI Chat Completions and Responses interfaces, the Anthropic Messages API and DashScope. It supports streaming, parallel tool calling, function calling, web search, structured output via JSON Schema, and explicit and implicit caching. For explicit caching, the docs list time-to-live options of five minutes and one hour. According to the documentation, batch inference, fine-tuning and custom deployments are not available.
Besides the API, Qwen4-Max is available on QwenCloud through a “Token Plan” and in Alibaba’s apps QwenWork and QwenStudio. The weights of the large model are not open. Alibaba says smaller models from the Qwen4 series will “soon” be released as open-weight models, but it gives no date.
Use cases from the announcement
Alibaba describes several scenarios, all based on its own tests or showcases. In cloud resource scheduling, the model is said to optimise its own workflows in a closed loop. Alibaba reports that this cut the cost of model calls for data annotation by 9 percent on average and made offline CPU rescheduling more than 14 times faster.
In an automotive manufacturing showcase, Qwen4-Max reportedly cut manual programming time for a workpiece from roughly 475 minutes to 24 minutes. In research, a 48-hour agent run over four iterations is said to have found an experimentally validated coolant molecule for chips. In an Alzheimer’s research experiment, Alibaba says the model executed 454 tool calls and 3,239 metric evaluations to design and optimise 24 candidate molecules. None of these examples has been independently verified.
Safety and data protection
On safety, Alibaba says software engineering training included code security checks for issues such as injection, memory safety and broken access control. For cybersecurity, training data was filtered by intent: material for finding and fixing vulnerabilities was kept, while offensive attack content was removed. Dedicated safety datasets are also meant to strengthen protections against chemical, biological, radiological and nuclear risks.
For companies in Europe, data processing is an important question. The documentation distinguishes between the deployment scopes “Global”, “International” and “Chinese mainland” across regions. If you plan to process personal or confidential data, clarify what the chosen region and deployment scope mean for storage and processing before you start, and review the contract terms.
What it means for developers
Qwen4-Max arrives at a time when providers compete mainly on price and efficiency for agents. Google recently introduced Gemini 4 Hydrogen, a fast model for agents, and a lot of money is flowing into agent platforms, as LangChain’s Series C shows. Qwen models are also already used in other products, for example in Apple Intelligence in China.
If you want to try Qwen4-Max, the OpenAI- and Anthropic-compatible interfaces often mean you only need to swap the base URL, API key and model ID. A comparison with real tasks from your own project makes sense, measuring not only quality but also the number of output and thinking tokens. Only that shows the actual cost per task.




