The numbers the preview withheld

On November 11, Alibaba opened a preview of Qwen4-Max and published almost nothing measurable: no parameter count, no benchmarks, no per-token price. That changed on November 15. The full launch post sets out the architecture and the evaluation tables, and QwenCloud's model page now carries the price list. It is the tables, not the model size, that carry the argument.

The backbone is 2.4 trillion parameters, the same total as Qwen3.8-Max. What shrank is the active footprint: 71 billion parameters per token, 25% fewer than the previous generation activates, with a further 328 billion parameters held in N-gram embeddings. Three Gated DeltaNet layers compress history into fixed-size states ahead of each sparse-attention layer, which is how Alibaba proposes to keep a 1-million-token context affordable. Whether that trade holds up in outside hands is an open question. The shape of it is clear enough: same warehouse, fewer lights on at once.

Where the 62% comes from

The headline figure is a cost per task, not a price cut. On Work360, Alibaba's own evaluation of professional work, Alibaba reports that the average task cost fell from $4.43 with Qwen3.8-Max to $1.67 with Qwen4-Max. Two things produce that drop. The model uses fewer tokens on the same work: about 24% fewer at Xhigh effort and 35% fewer at High effort on Work360, and roughly half as many on Terminal-Bench 4.0. And the tokens themselves are cheaper: $1.50 per million in, $4.50 per million out on QwenCloud, with implicit caching at $0.075 per million.

Read that carefully. A cost-per-task figure on the vendor's own benchmark bundles together the vendor's prices, the vendor's task mix and the vendor's scoring. The 62% is internally consistent arithmetic, 62.3% on the rounded endpoints Alibaba published. It is not a quote for your workload, and no independent evaluation has reproduced it yet. Teams running long agent jobs should treat it as a reason to run their own test this week, not as a budget line.

The chart Alibaba published against itself

The most useful page of the launch is the comparison chart, because it does not flatter the new model everywhere. On the HLAB-Vals-1.1 legal-work test, Qwen4-Max scored 21.3, well ahead of Anthropic's Opus 5.5 at 9.2 and OpenAI's GPT-6 Astra at 7.1. On Terminal-Bench 4.0, a widely used external test of terminal-based agent work, Qwen4-Max scored 48.5, behind Opus 5.5 at 59.6 and Kimi K3.1 at 67.2, on the scores Alibaba itself published. The launch leads with the test where it wins by the widest margin and still prints the one where it trails. A launch that does that is giving readers the material to discount it properly.

The Work360 score itself rose from 47.74 to 52.69. That is a solid gain on a home benchmark, and it should be weighted accordingly.

The demos are showcases, not evidence

The launch post leans on long-horizon demonstrations: a 48-hour continuous run that built a candidate space of 2.7 million molecules and, through four closed design-and-test iterations, obtained an experimentally validated coolant molecule; a drug-discovery run against the NLRP3 inflammasome ending in a candidate with an estimated binding affinity about 3.6 times the reference; a manufacturing case where programming one workpiece reportedly fell from about 475 minutes of manual effort to 24. These are individual showcase cases, several with partner platforms, and the drug candidate still awaits wet-lab validation. They show that an agent loop of this length can finish. They say little about how often it finishes on work Alibaba did not choose.

The quieter numbers may matter more. On Alibaba's Secure Code Bench, code passing both functionality and security checks at function level rose from 22.0% to 31.8%, and code secure at function, file and repository level together rose from 8% to 14.7%. Progress, stated plainly: most evaluated outputs still did not pass all of the stated functionality and security criteria.

What is actually available today

The API is live on QwenCloud, through the Token Plan or pay-as-you-go, with OpenAI- and Anthropic-compatible interfaces, and the model is built into Qwen Studio, QwenWork and Qoder. The weights are not open. Alibaba says smaller Qwen4 models will be released as open weights, in the same family, with no date attached. Anyone whose interest in Qwen4 is self-hosting is still waiting, and the announcement gives them nothing to plan against beyond the promise.

Sources

  1. [1] Qwen Team, "Qwen4-Max: Your New Teammate Has Arrived," Qwen Blog, November 15, 2026Read source
  2. [2] QwenCloud, Qwen4-Max model page (pricing as listed, updated November 15, 2026)Read source
  3. [3] DEV Community analysis of the launch post, November 16, 2026Read source
  4. [4] AI Letter (ailetter.ai.kr) launch report, published 2026-11-15T23:09:26ZRead source