Alibaba unveiled Qwen4-Max on Sunday, a new flagship AI model the company says handles complex professional and scientific work with fewer tokens and at lower cost than its predecessor, Qwen3.8-Max, which launched in August. On Alibaba's own Work360 benchmark, cost per task fell 62%, from $4.43 to $1.67.
Qwen4-Max is a mixture-of-experts model with 2.4 trillion total parameters, 71 billion of which are active for each token, plus 328 billion parameters for a new N-gram embedding. It is a multimodal reasoning model that accepts text and images, with a context window of 1 million tokens.
The efficiency gains come from token usage. Alibaba says Qwen4-Max uses up to 35% fewer tokens than Qwen3.8-Max on Work360 and 49% fewer on Terminal-Bench 4.0.
Key facts
- 01Alibaba unveiled Qwen4-Max on Nov. 15, 2026, replacing Qwen3.8-Max, which launched in August.
- 02Cost per task on Alibaba's Work360 benchmark fell 62%, from $4.43 to $1.67.
- 03The model has 2.4 trillion parameters, 71 billion active per token, and a 1 million-token context window.
- 04Qwen4-Max scores 43 on the Artificial Analysis Intelligence Index, versus 65 for Claude Fable 5.2.
- 05API pricing is $1.50 per million input tokens and $4.50 per million output tokens, per Artificial Analysis.
In Alibaba's published benchmark table, Qwen4-Max posts its clearest lead in legal agent work. It scored 21.3 on HLAB-Vals 1.1, against 9.2 for Anthropic's Claude Opus 5.5, 6.3 for OpenAI's GPT-6.1 Sol, 7.1 for GPT-6 Astra and 11.7 for Moonshot AI's Kimi K3.1. On the DeepSWE v1.1 coding benchmark, it scored 75.7, narrowly ahead of Kimi K3.1 at 75.5, GPT-6.1 Sol at 75.2 and Claude Opus 5.5 at 74.2.
The same table shows Qwen4-Max trailing on agentic coding. It scored 48.5 on Terminal-Bench 4.0, versus 67.2 for Kimi K3.1 and 59.6 for Claude Opus 5.5, and 31.5 on ProgramBench, versus 65.0 for Claude Opus 5.5. On Humanity's Last Exam, it scored 51.7 against 59.8 for Claude Opus 5.5.
Independent testing puts the model in the middle of the frontier pack. Qwen4-Max scored 43 on the Artificial Analysis Intelligence Index, behind Claude Fable 5.2 at 65, Gemini 4 Carbon at 60, Claude Opus 5.5 at 59, GPT-6 Astra at 57 and Kimi K3.1 Pro at 46, and ahead of xAI's Grok 4.7 at 35. Among 81 models in its class, it ranked 38th, against a median score of 41.
Price is where Alibaba competes hardest. Artificial Analysis lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output tokens, with a 95% discount on cached inputs. Its average task cost on the Intelligence Index was $3.27, compared with $10.04 for Claude Opus 5.5; only GPT-6.1 Sol was cheaper, at $1.27.
Alibaba also published customer examples. At two automotive suppliers, programming effort for a single workpiece fell from about 475 minutes to 24. In a legal due-diligence task, the model sorted 198 working papers into an 86-page report, and in a 48-hour autonomous run it designed a data-center coolant that an automated lab then synthesized.
Alibaba paired the launch with safety changes. Training included security checks alongside coding tests to reduce vulnerabilities such as code injection, memory-safety flaws and broken access controls, and the company strengthened safeguards against misuse in cybercrime and hazardous materials. Alibaba said internal testing found no evidence of significant high-risk capabilities.
Qwen4-Max is available through the QwenCloud API and in the QwenWork, Qoder and QwenStudio apps, along with a new Live Delivery feature that ships an early working version of a project and iterates on user feedback. The model's weights are not downloadable, though Alibaba says smaller Qwen4 models will be released as open source. The company has not disclosed what hardware it used to train the model.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




