·4 min read
Qwen4-Max: Multimodal Capabilities, Pricing, and AI Agent Evaluation

Alibaba Cloud announced Qwen4-Max on November 16, 2026, positioning it for coding, multimodal understanding, and AI agent tasks. Its documented model ID is qwen4-max. Teams can first check regional availability, pricing, and supported features in Model Studio, then evaluate it on their own data and tasks. Official announcement
What can Qwen4-Max do?
Alibaba says the mixture-of-experts model has 2.4 trillion total parameters, activates 71 billion per token, and has a one-million-token context window. Total and activated parameters describe different things; a large context window also does not guarantee accurate retrieval of every detail.
The model documentation lists text, image, and video inputs with text output, plus tool calling, web search, and structured output. Potential uses include chart interpretation, document analysis, coding, and multi-step tasks. Validate the results for your own needs. Visual understanding does not mean image or video generation.
How should teams read the benchmark claims?
Alibaba reports that, on its proprietary Work360 benchmark, Qwen4-Max outperformed Qwen3.8-Max overall while using 35% fewer tokens at High effort and reducing average cost per task by about 60%. These are vendor-reported results under specific conditions, not guaranteed savings for all workloads. Actual cost depends on inputs, outputs, caching, retries, and tools.
How does the released model differ from the preview?
The released model ID is qwen4-max; qwen4-max-preview is separate. The official preview documentation says it is available only through Token Plan, with the preview ending at 24:00 on November 25, 2026 (UTC+8), after which it will be retired. Preview users should check regional availability, parameters, and billing for the released model and run regression tests before switching. Preview documentation
For the preview’s OpenAI Chat Completions-compatible interface, the documentation says max_tokens and max_completion_tokens limit thinking and final-answer tokens together. Check the target model and interface documentation during migration so an undersized output budget does not truncate responses; do not assume defaults are identical across versions.
Pricing and regional availability
Alibaba Cloud lists the Singapore international deployment of qwen4-max at $1.50 per million input tokens and $4.50 per million output tokens for up to one million input tokens per request. Output billing includes thinking and the answer. Prices differ across regions and deployment scopes; caching, tools, plans, and taxes can also affect a bill. Check the current official regional pricing.
At those Singapore rates, a task using 900,000 input tokens and 100,000 output tokens has a model-token cost of $1.35 + $0.45 = $1.80. This is an arithmetic example, excluding caching and other services, not a measured bill.
How to evaluate the model
- Confirm Model Studio account, region, and eligibility. Follow the official documentation for the API key and endpoint, using model ID qwen4-max.
- Choose representative tasks, such as chart-based document questions, existing-code fixes, and multi-step tool calls, and compare against your current model under the same conditions.
- Measure completion rate, human corrections, latency, input and output tokens, and full task cost. For long documents, check citations and omissions separately.
- Begin in a test environment with restricted tool permissions and human review, then expand gradually. API-format compatibility does not imply identical parameters, defaults, or error handling.
Some capabilities vary by region, including batch inference support. Verify the selected region’s documentation. We did not call the model or test code for this article. Regional capabilities
Frequently asked questions
Are Qwen4-Max weights open?
The announcement confirms service access to Qwen4-Max and says smaller open-weight Qwen4 models are planned. That does not establish that Qwen4-Max weights are open.
Can it replace an existing API model directly?
The documentation lists compatibility with some OpenAI and Anthropic API formats, but endpoints, authentication, parameters, regions, and tool behavior still require checking and regression tests.
Is it suitable for AI agents?
Tool calling, search, and structured output make it a candidate for evaluation. Actual success and safety depend on the task, permissions, and acceptance process.
Official sources
- Alibaba Cloud announcement, November 16, 2026.
- Qwen4-Max model and regional documentation.
- Qwen4-Max-Preview documentation.
Verified November 17, 2026. Pricing, availability, and capabilities can change.