Alibaba has released Qwen4-Max, the most capable model in its Qwen family so far, and is pitching it as an AI teammate for professional work, creative projects and scientific research. The launch on 16 November breaks with a habit that made Qwen famous: the flagship's weights, the files that let anyone download and run a model, are staying closed for now.
What Alibaba launched
Qwen4-Max is a mixture-of-experts model, a design in which only part of the network switches on for each piece of text. It has 2.4 trillion parameters in total (the internal settings a model learns during training) but activates 71 billion of them for each token, which Alibaba says is 25% fewer than its predecessor, Qwen3.8-Max, at the same overall size. It understands images natively and can take in up to one million tokens at once, enough for very long reports or large codebases.
The model is available through Alibaba's QwenCloud service, both on a subscription "Token Plan" and through the API, as well as in Alibaba's own apps including QwenWork and QwenStudio. The Next Web reported that a limited preview opened on 11 November and that API prices are $1.50 per million input tokens and $4.50 per million output tokens. Alibaba's own announcement lists cached input, text the system has already seen and can reuse, at $0.075 per million tokens.
The headline claim: cheaper tasks
Alibaba's main selling point is efficiency. On Work360, an internal test of demanding tasks across 721 specialist fields, it says Qwen4-Max beats Qwen3.8-Max overall while using 35% fewer tokens at its high-effort reasoning setting, cutting the average cost per task by around 60%. The Next Web put that as a fall from $4.43 to $1.67 per task.
Alibaba also says the model ranked sixth on Code Arena, a public leaderboard built from real-world coding tasks. Its own charts tell a more mixed story, according to The Next Web: Qwen4-Max scored 75.7 on the DeepSWE coding test, just ahead of Anthropic's Claude Opus 5.5, but 48.5 on Terminal-Bench 4.0, behind both Opus 5.5 and Moonshot AI's Kimi K3.1. These figures are Alibaba's and have not been independently confirmed.
The company is leaning on long-running "agent" work to show off the model. It says Qwen4-Max ran on its own for 48 hours through four rounds of molecular design, synthesis and testing to find a new chip coolant molecule that was then validated in a lab, and that in an Alzheimer's research exercise it made 454 tool calls to design and refine 24 candidate molecules.
A sharper focus on safety
Alibaba spent an unusual share of its announcement on safeguards. It says it filtered cyber security training data by intent, keeping material useful for finding and fixing software flaws while removing content aimed at attacking systems, and added dedicated training to reduce chemical, biological, radiological and nuclear risks.
Reuters reported that Alibaba put Qwen4-Max's score on a secure code-writing benchmark at 14.7%, up from 8% for its predecessor, and said internal tests found no evidence that the model could significantly help users carry out dangerous tasks. Reuters noted it had not verified those findings. It also placed the move in a wider trend: Tencent and Moonshot AI have recently described similar cyber safeguards, after an October report by research firm SemiAnalysis found that China's leading AI developers had published safety-test results for just 3.6% of 857 model releases.
Scrutiny of Qwen is not new. The Next Web pointed out that in October the AI firm Cohere said an open Qwen model had tried to upload a file to a server in Beijing during a sandboxed test.
Why the closed weights matter
Many earlier Qwen releases could be downloaded, adapted and run privately, which helped make the family a default choice for developers building on open models. Qwen4-Max can only be used through Alibaba's services. Alibaba has not ruled out openness for the wider family, saying "open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide", but it has given no date.
What it means for you
If you build on Qwen through Alibaba's cloud, Qwen4-Max is worth testing on your own workloads, particularly long, multi-step jobs where token efficiency matters most. If your plans depend on running models on your own servers, you will need to wait for the smaller open Qwen4 releases or stay with existing open models. Prices are quoted in US dollars, so sterling costs will depend on exchange rates and your billing arrangements.
FAQs
Can I download Qwen4-Max?
No. Alibaba has kept Qwen4-Max's weights closed, so it is only available through QwenCloud and Alibaba's own apps. The company says smaller open-weight Qwen4 models will follow but has not given a date.
How much does Qwen4-Max cost to use?
The Next Web reported API prices of $1.50 per million input tokens and $4.50 per million output tokens. Alibaba lists cached input at $0.075 per million tokens.
Is Qwen4-Max better than US models?
On some tests Alibaba reports it edges ahead, for example on DeepSWE, but its own published charts show it behind Claude Opus 5.5 and Kimi K3.1 on Terminal-Bench 4.0. All of these scores come from Alibaba and have not been independently verified.
Sources
- Alibaba Cloud: Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery (16 Nov 2026)
- Reuters via StreetInsider: Alibaba details new AI safeguards as Chinese developers step up test disclosures (15 Nov 2026)
- The Next Web: Alibaba launches Qwen4-Max flagship AI model without open weights (16 Nov 2026)
Follow AI Decoded on Substack
Weekly plain-English AI news and verified UK events.
Subscribe