Alibaba released Qwen4-Max on November 16, 2026, bringing its new flagship multimodal AI model to QwenCloud through the Token Plan and API, as well as QwenWork and QwenStudio. Alibaba describes a sparse architecture with 2.4 trillion total parameters, 71 billion activated per token, and native image and video understanding. The announcement also specifies a one-million-token context window.
This is the production release, separate from the earlier preview. For developers, the most useful details sit in the technical documentation: the exact model ID, regional feature differences, reasoning defaults and a cached-input price that needs careful interpretation.
What Changed From the Preview
Alibaba’s model release changelog records two distinct releases:
- November 11:
qwen4-max-preview, available through Token Plan only. - November 16:
qwen4-max, listed in Singapore with an International service deployment scope.
Applications targeting the production model should use qwen4-max as the model parameter. The preview name is a separate identifier, so developers should check existing configurations instead of assuming the production release automatically changes what their application calls.
Alibaba lists three access routes in its announcement. QwenCloud provides Token Plan and API access; QwenWork and QwenStudio offer access within Alibaba’s work products. The announcement confirms availability across those products without detailing an identical feature set or subscription entitlement for each.
For an application integration, I’d start with the API documentation. Product availability answers where people can reach the model; the model card answers which inputs, outputs and controls an application can actually use.
A 2.4-Trillion-Parameter Model With Sparse Activation
Alibaba says Qwen4-Max uses a mixture-of-experts, or MoE, architecture. This approach activates a subset of the model’s parameters for each token instead of using every parameter at every step.
| Specification | Published detail |
|---|---|
| Total parameters | 2.4 trillion |
| Activated parameters per token | 71 billion |
| Input modalities | Text, images and video |
| Output modality | Text |
| Context window | 1,000,000 tokens |
| Fine-tuning | Not supported |
| Model deployment option | Not supported in the documented service |
The architecture figures come from Alibaba’s announcement; input, output and service capabilities are listed in the model documentation.
Dividing 71 billion by 2.4 trillion puts the activated share at roughly 3% of total parameters per token. That helps explain the design, but it is not a claim that inference costs only 3% as much as a dense model. Attention, memory movement, routing and serving infrastructure also affect cost and speed.
Alibaba reports 25% fewer activated parameters than Qwen3.8-Max at the same 2.4-trillion-parameter scale. That is an architectural comparison, not an independently measured latency result.
Native vision also needs a precise reading: Qwen4-Max accepts images and video, but its documented output is text. Developers can build visual analysis workflows around it; this model listing does not provide image or video generation.
The Million-Token Window Has Separate Limits
The headline context capacity is one million tokens, but the model card lists different input and output ceilings depending on whether thinking mode is enabled.
| Documented limit | Non-thinking mode | Thinking mode |
|---|---|---|
| Maximum input | 991,808 tokens | 983,616 tokens |
| Maximum output | 131,072 tokens | 262,144 tokens |
| Overall context length | 1,000,000 tokens | 1,000,000 tokens |
These are separate ceilings, not amounts developers should add together and expect to use simultaneously. A request still needs to fit the overall context constraints.
That distinction matters for large codebases, document collections and long-running agents. An application that fills nearly the entire window with input cannot safely assume it also has the maximum advertised output allocation available.
Tokens are model-processing units, not words, pages or seconds of video. The published limits therefore do not translate into a fixed document count or video duration.
My view is that the exact input limits are more useful for implementation than the rounded one-million figure. Teams should budget headroom for instructions, tool results and responses, then evaluate retrieval accuracy on their own material. A large context window specifies capacity; it does not tell you how reliably a model will find every relevant detail inside it.
API Compatibility Comes With Reasoning Defaults
Qwen4-Max supports OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and Alibaba’s DashScope interfaces. Its documented capabilities include streaming, web search, parallel tool calls, function calling, and structured output through JSON Object and JSON Schema modes. Explicit and implicit caching are also supported. Alibaba’s model card lists these features across its documented regions.
For OpenAI Chat Completions, the model-selection and reasoning fields can look like this:
{
"model": "qwen4-max",
"reasoning_effort": "low"
}
This is a configuration fragment, not a complete API request. An integration still needs authentication, the appropriate endpoint and its message payload. The fragment selects the production model and requests the low reasoning-effort setting.
The documented effort values are low, high and xhigh. Thinking is enabled by default. When neither reasoning effort nor thinking budget is supplied, Alibaba specifies high effort with a 131,072-token thinking budget.
That budget is an allowance, not evidence that every request consumes that many reasoning tokens. Still, leaving it implicit makes cost and response-time comparisons harder to interpret. I’d set reasoning effort explicitly in an evaluation so a model change does not quietly become a configuration change too.
Regional differences also deserve attention. The documentation lists batch inference as supported in Beijing, but unsupported in Singapore, Frankfurt, Virginia, Tokyo and Hong Kong. A workflow built around offline batches therefore needs a regional capability check before adoption.
The $0.075 Price Applies to Cached Input
Alibaba’s announcement gives a price of $0.075 per million input tokens for an implicit cache hit. That condition is essential: it is not a blanket price for all input tokens, and it is not an output-token price. The press release labels it specifically as “Input (Implicit Cache Hit).”
At that rate, one million qualifying cached input tokens would cost 7.5 cents. The figure alone cannot price an entire request containing uncached input and generated output.
A practical cost estimate needs to separate:
- Input tokens billed at the applicable uncached rate.
- Input tokens qualifying for a cache-hit rate.
Frequently Asked Questions
3 questions
1Which Model ID Should Developers Use?
The production model ID is qwen4-max. Alibaba separately listed qwen4-max-preview for the earlier Token Plan-only preview.