Qwen4-Max Runs 71 Billion of Its 2.4 Trillion Parameters per Token and Opens at $1.50 per Million Input Tokens

Alibaba released Qwen4-Max on November 16, 2026, a 2.4-trillion-parameter mixture-of-experts model with native image and video understanding and a one-million-token context window. It is API-only on QwenCloud and Model Studio at $1.50 input and $4.50 output per million tokens internationally; Alibaba says smaller Qwen4 models will get open weights.

A librarian on a rolling ladder reaches into a small glowing cluster of open drawers on a vast cobalt wall of closed drawers.
A librarian on a rolling ladder reaches into a small glowing cluster of open drawers on a vast cobalt wall of closed drawers.

Alibaba’s new flagship keeps its predecessor’s total size but switches on a quarter fewer parameters per token. Qwen4-Max, released on November 16, 2026, is a mixture-of-experts (MoE) model with 2.4 trillion total parameters that activates 71 billion for each token. Alibaba says that is 25% fewer active parameters than Qwen3.8-Max at the same 2.4-trillion scale, and it credits the new Qwen4 architecture, which revises the sparse MoE design across attention, residual connections and embeddings.

The model accepts text, images and video and produces text, with a one-million-token context window. Alibaba positions it for coding, office work, front-end design and long-running agent tasks.

QwenCloud, Model Studio and per-region token prices

Qwen4-Max is available through QwenCloud’s Token Plan and API, and inside Alibaba’s own apps, including QwenWork and QwenStudio. The Model Studio release log lists qwen4-max on November 16 for both the International and Global deployment scopes. A qwen4-max-preview model had appeared on November 11, available only through the Token Plan.

Alibaba’s documentation lists pricing for the Singapore region’s International scope at $1.50 per million input tokens and $4.50 per million output tokens, the same in thinking and non-thinking modes for requests up to one million tokens. Several other regions, including Beijing, Hong Kong, Frankfurt and Virginia, list $1.413 input and $4.24 output. Alibaba’s launch post puts the implicit cache-hit input price at $0.075 per million tokens. The Singapore listing also includes a free quota of one million tokens, valid for 90 days.

The model page details limits. Maximum output is 131,072 tokens in non-thinking mode, and thinking plus the final answer can total 262,144 tokens. Thinking is on by default with reasoning_effort set to high; the supported levels are low, high and xhigh, and the docs show how OpenAI and Anthropic effort values map onto them. The model works with OpenAI’s Chat Completions and Responses APIs, Anthropic’s Messages API and Alibaba’s DashScope. Web search, parallel tool calling, function calling, structured JSON output, prefix completion and explicit and implicit caching are supported. Batch inference, fine-tuning and self-deployment are listed as not supported.

No open weights for the flagship

Unlike many Qwen releases, Qwen4-Max is not downloadable. Alibaba’s announcement says only that “open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide,” without naming sizes, licences or dates. For developers who build on Qwen checkpoints, the flagship gives an early look at the Qwen4 architecture, but working weights will have to wait for the smaller models.

The benchmark claims

Alibaba says Qwen4-Max ranked 6th on Code Arena and scored well on Agents’ Last Exam, Vals Index, JobBench and Chartography, though the launch post does not list the scores.

Most of its efficiency claims rest on Work360, an Alibaba benchmark of difficult industry tasks spanning 721 specialised fields. On Work360, the company says Qwen4-Max beats Qwen3.8-Max overall while using 35% fewer tokens at High reasoning effort and cutting average cost per task by around 60%. Work360 is proprietary, so those numbers cannot be checked independently.

Machining, chip-coolant and Alzheimer’s demos

The rest of the announcement is a series of company demonstrations. In cloud infrastructure, Alibaba says the model tuned its own scheduling workflows, cutting data-annotation model call costs by an average of 9% and speeding up offline CPU rescheduling by more than 14 times. In an automotive manufacturing demo, Alibaba says it cut manual workpiece programming from about 475 minutes to 24, catching overcuts and under-removal before machining.

For science, Alibaba describes a continuous 48-hour agent run through four rounds of molecular design, synthesis, testing and active learning that produced an experimentally validated coolant molecule for chip cooling. In an Alzheimer’s research exercise, the model carried out 454 tool calls and 3,239 metric evaluations to design and refine 24 candidate molecules.

Code-security and CBRN filtering in training

Alibaba says code-security checks ran alongside functional tests during software-engineering training, aiming to reduce injection, memory-safety and access-control flaws in generated code. For cybersecurity, it filtered training data by intent, keeping defensive material on finding and fixing vulnerabilities and removing offensive attack content. It also cites dedicated datasets and post-training aimed at chemical, biological, radiological and nuclear risks.

For now, Qwen4-Max is a hosted model priced to compete on cost per task. How the Qwen4 design performs outside Alibaba’s own benchmarks will become clearer when the open-weight models arrive and can be tested directly.

Topics
  • #Alibaba
  • #Qwen
  • #Qwen4-Max
  • #Mixture of Experts
  • #Model APIs
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.