Alibaba's Qwen4-Max is out: 1 million tokens of context, $1.50 per million input tokens, and no downloadable weights
Quick answer: Alibaba released Qwen4-Max on Nov 16, 2026 (Hangzhou time), its most capable Qwen model so far. It has 2.4 trillion total parameters, reads text, images and video, and takes up to a 1-million-token context. Through the QwenCloud API it costs $1.50 per million input tokens and $4.50 per million output tokens. Unlike many earlier Qwen releases, you can’t download it. Alibaba says smaller open-weight Qwen4 models are coming, with no date. (Alibaba Cloud, Nov 16; QwenCloud model page; TNW, Nov 16)

The specs and price
- Size: a mixture-of-experts model, so only part runs for each request: 2.4 trillion parameters in total, 71 billion active per token. (Alibaba)
- Context: 1 million tokens; the model page lists a 991K maximum input. (QwenCloud)
- Price per million tokens: $1.50 input, $4.50 output, and $0.075 for input that hits the cache. (QwenCloud)
- Where: the QwenCloud API and Token Plan, plus Alibaba’s QwenWork and QwenStudio apps. A limited preview opened Nov 11. (Alibaba; TNW)
What Alibaba claims (its own numbers)
- On Work360, Alibaba’s own benchmark of tasks across 721 specialized fields, Qwen4-Max beats the earlier Qwen3.8-Max overall while using 35% fewer tokens at high effort and cutting the average cost per task by about 60%. TNW reports the average cost per task falling from $4.43 to $1.67. (Alibaba, Nov 16; TNW, Nov 16)
- It ranks 6th on Code Arena, a real-world coding leaderboard, according to Alibaba’s Nov 16 launch post. (Alibaba)
- Alibaba says the model ran on its own for 48 hours to design a chip-cooling molecule that was then made and tested in a lab. (Alibaba)
The mixed picture
TNW says Alibaba’s own launch charts (Nov 16), as read by FourWeekMBA, are more mixed than the headline: Qwen4-Max scored 75.7 on the DeepSWE v1.1 coding test, just ahead of Claude Opus 5.5 at 74.2, but 48.5 on Terminal-Bench 4.0, behind Opus 5.5 at 59.6 and Kimi K3.1 at 67.2. Those are benchmark scores where higher is better. We have not seen independent testing of any of these results, and Work360 is Alibaba’s own test.
What it means if you build with it
- Closed weights: you can use it only through Alibaba’s service. Smaller open-weight Qwen4 models are promised “soon,” without a date. (Alibaba)
- Safety framing: Alibaba says it filtered training data to keep defensive cyber material and remove offensive attack content, and added safety training for chemical, biological, radiological and nuclear risks. That is a description of its process, not an outside audit. (Alibaba)
- Test on your own work. The cost and token claims come from Alibaba’s benchmark; your results will depend on your prompts.
FAQ
Can I run Qwen4-Max myself? No. The weights are not released.
Is it free? The sources list per-token API prices. We did not see a free tier.
Is it better than Claude or Gemini? Alibaba says it is near the top on some coding and agent tests. Independent comparisons are not in yet.
Sources
- Alibaba Cloud Community, “Alibaba Unveils Qwen4-Max” (Nov 16, 2026): https://www.alibabacloud.com/blog/alibaba-unveils-qwen4-max-new-ai-teammate-for-work-creativity-and-discovery_603783
- QwenCloud, Qwen4-Max model page (pricing, limits): https://www.qwencloud.com/models/qwen4-max
- The Next Web, “Alibaba launches Qwen4-Max flagship AI model without open weights” (Nov 16, 2026): https://thenextweb.com/news/alibaba-launches-qwen4-max-flagship-ai-model-without-open-weights
— AltTab Dad
