HomeAIAI ModelsAlibaba Releases Qwen4-Max With a One-Million-Token Window

Alibaba Releases Qwen4-Max With a One-Million-Token Window

Alibaba has released Qwen4-Max, which it calls the most capable model in the Qwen family to date. In a 16 November blog post, Alibaba Cloud said the model has “2.4 trillion total parameters, native vision capabilities, and a one-million-token context window.” Alibaba Cloud’s Model Studio documentation, last updated on 16 November 2026, lists the invocation ID as qwen4-max and gives the same 1,000,000-token context length.

The release follows Alibaba’s earlier opening of Qwen4-Max-Preview on Token Plan only, which TSN covered as a preview rather than a full launch. The new post says Qwen4-Max is available now on QwenCloud via the Token Plan and API, and on QwenWork and QwenStudio.

Alibaba says the model activates 71 billion parameters per token

The blog describes Qwen4-Max as a sparse mixture-of-experts model. It “activates only 71 billion per token—25% fewer activated parameters compared to Qwen3.8-Max” at the same 2.4-trillion-parameter scale, Alibaba said, “delivering stronger capabilities at a lower inference cost.” Those efficiency and quality claims are the company’s.

So are the benchmark scores. Alibaba said Qwen4-Max “ranked 6th on Code Arena, a real-world coding leaderboard,” and posted strong scores on Agents’ Last Exam, Vals Index, JobBench and Chartography. On Work360, which it describes as a proprietary benchmark covering 721 specialised fields, it said Qwen4-Max outperforms Qwen3.8-Max overall “while using 35% fewer tokens at High effort reasoning mode and reducing the average cost per task by around 60%.”

The only unit price stated in the announcement is for input with an implicit cache hit: “$0.075 per 1M tokens.” The Model Studio page points to a separate pricing document for prices and free quotas, and it lists compatibility with the OpenAI Chat Completions and Responses APIs, the Anthropic Messages API and DashScope. Input modalities are text, image and video; output is text. The same tables list streaming, deep thinking, function calling, web search, structured output and both explicit and implicit prompt cache as supported. Batch inference, fine-tuning and model deployment are listed as not supported. The documentation shows the model in several Model Studio regions, including Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong. Default thinking effort is set to high. Alibaba also said “open-weight, smaller-scale models from the Qwen4 series will soon be released.” No date is given.

That promised weight release, if it comes, would sit alongside earlier third-party work on Qwen open weights, including Nous’s Arete-27B adaptation. It is a plan, not a downloadable model today.

A company scorecard is not an independent ranking

The cleanest way to read the 16 November announcement is as a full product launch with a company-run scorecard attached. The preview on Token Plan is no longer the whole story: the model ID qwen4-max is now documented on Model Studio, with a one-million-token context window and a list of supported APIs. That is a genuine change from the preview TSN already recorded. It is also as far as the official pages go on independently checked quality.

Work360 is Alibaba’s own benchmark. The 35 per cent token saving and the “around 60%” cut in average cost per task are measured there, at High effort reasoning mode, against Qwen3.8-Max. Code Arena’s sixth place is likewise reported by Alibaba. None of those figures is an external audit, and none of them tells a buyer what a given task will cost once the implicit-cache-hit rate is lower than the blog’s headline price. The only unit price Alibaba printed in the launch post is $0.075 per 1 million input tokens on an implicit cache hit.

The same caution applies to the open-weight line. Alibaba said smaller Qwen4 models will soon be released. Until they are, this launch is an API and platform opening, not a weight drop. A previous TSN report on a Cohere claim involving a Qwen model is a reminder that questions about how Qwen models behave in the wild sit on a separate track from a vendor’s launch benchmarks. Alibaba’s blog also describes its own safety training and data-filtering steps; those, too, are company statements.

The next official facts worth watching are ordinary ones: the rest of the price list, and whether the promised smaller open-weight Qwen4 models actually appear. Until then, the record is a documented model ID, a one-million-token window, and a set of Alibaba-stated scores.

Sources

Related stories

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.