Alibaba launches Qwen4-Max: new top model reads a million tokens in one go
Alibaba Cloud has released Qwen4-Max, its new top model that understands text, images and video and can take in a million tokens at once. Anyone using it in the Frankfurt region pays about 14 per cent less than for its predecessor Qwen3.8-Max.
, Alibaba Cloud released Qwen4-Max, the new top model of its Qwen family. Developers can use it through Model Studio, Alibaba's cloud service where companies rent AI models. The model is available in, among other regions, Frankfurt.
In a single request, the model can take in up to one million tokens: the small pieces of text, often a short word or part of a word, in which an AI model reads and writes. This lets it read a long document or a lot of program code in one go. Qwen4-Max also understands images and video and answers in text.
Alibaba intends the model for programming, office work, designing websites and apps, and complex tasks for AI agents: programs that carry out a series of steps by themselves, such as looking something up, writing code and testing it. For this, the model can search the web and call other programs.
Qwen4-Max also works with the APIs of OpenAI and Anthropic, the standard connections through which programs call those companies' models. Developers who already build with them can call Alibaba's model in the same way. Independent tests will show how the model compares with other top models.
Cheaper than its predecessor
In the Frankfurt region, Qwen4-Max costs $1.41 per million tokens that a developer sends in (about €1.26) and $4.24 per million tokens that the model writes back (about €3.79). For its predecessor Qwen3.8-Max in Frankfurt, this was $1.65 and $4.95, so the new model is about 14 per cent cheaper.
Input that keeps coming back, such as fixed instructions or a document an agent keeps referring to, can be stored by the model for a while. Storing it costs a little extra, but reusing the stored input afterwards costs $0.071 per million tokens in Frankfurt: a twentieth of the normal price.
Think first, then answer
By default, Qwen4-Max thinks before it answers. Developers can choose how thoroughly: low, high (the default) or extra high. The thinking and the answer together must stay within 262,144 tokens. So the longer the model thinks, the less room is left for the answer.
A test version has been available since 11 November to people with an Alibaba subscription. That test version will be withdrawn on 25 November 2026.
What does this mean for the road to the singularity?
AI agents that program or work through documents on their own already exist. At every step they read the same instructions and the same code again, and it is exactly this reading that makes long tasks expensive. Qwen4-Max makes that reading cheaper, and rereading stored input even twenty times cheaper. Because the model uses the same connection as the models of OpenAI and Anthropic, a developer can also switch between those models more easily. When switching is this simple, customers pick, task by task, the model that does good work for the least money. A more expensive model then has to be clearly better, and so every price cut by one company pulls the prices of the others down with it.
If that trend continues, it becomes affordable for a small company or a single programmer to let an agent work on a project for hours: updating old software, tracking down bugs, building a website. Work that now needs a team could then be done by one person with a few agents. And because AI companies themselves also do a great deal of programming, cheaper programming work by AI eventually also helps to build the next generation of AI.
What is the technological singularity? Read the explanation →