A closed slim grey laptop and a clear glass cup of green tea on a wooden desk lit warm by a small black desk lamp beside a potted plant, in front of a large window at night that shows the blurred lights of tall city buildings far away.

The Models

Alibaba Releases Qwen4-⁠Max, and Cuts the Price

Alibaba’s closed flagship, with a 2.4-trillion-parameter backbone, costs $1.50 per million input tokens, a quarter less than its predecessor; on its own test, tasks cost 62 percent less.

Alibaba released Qwen4-Max late Sunday, the largest model in its new Qwen4 family, and kept its weights to itself. Smaller Qwen4 models will follow as open source, the company said; it gave no date.

The model has a 2.4-trillion-parameter backbone, of which 71 billion parameters are activated for each token, plus a further 328 billion “N-gram Embedding” parameters, Alibaba’s Qwen team wrote in its announcement. Active ParametersQwen4-Max is a sparse mixture-of-experts model: each token passes through 71 billion of its 2.4 trillion backbone parameters, about 3 percent. It reads text, images and video, and holds a context of one million tokens. It is also cheaper to use than its predecessor, Qwen3.8-Max.

On QwenCloud, Alibaba’s developer platform, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens on the Singapore-region price list, against $2 and $6 for Qwen3.8-Max. Cached input costs 7.5 cents per million tokens, down from 25 cents, according to the price list on Monday. The South China Morning Post, which Alibaba owns, put the model’s input price at “about 8 US cents per 1 million tokens”; that matches the cached rate, not the price of fresh input.

The announcement came on Monday in China, late Sunday night in New York. The Post reported that Qwen4-Max is the same size as Qwen3.8-Max, a 2.4-trillion-parameter backbone, but activates 25 percent fewer per token, and that Alibaba had made the Max series open source for the first time with Qwen3.8, in August.

Its Own Measure

Alibaba says the average cost of a task on Work360, a benchmark it built itself, fell 62 percent, from $4.43 with Qwen3.8-Max to $1.67. The announcement also reports that the new model uses fewer tokens: on Work360, about 24 percent fewer at its “Xhigh” effort setting and 35 percent fewer at “High.” Its Work360 score rose to 52.69, from 47.74.

On a chart of 16 tests in Alibaba’s announcement, Qwen4-Max beat Anthropic’s Claude Opus 5.5 on six and trailed it on ten. It led on DeepSWE, a coding test, 75.7 to 74.2, and fell well short on Terminal-Bench 4.0, 48.5 to 59.6. The scores are as Alibaba reports them; several are cited from outside leaderboards.

Artificial Analysis, an independent firm that runs its own evaluations, scored Qwen4-Max 43 on its Intelligence Index, below Kimi K3.1 Pro at 46 and Anthropic’s Claude Fable 5.2 at 65. Those scores are the firm’s, not Alibaba’s, and were taken at different reasoning settings: extra-high for Qwen4-Max, maximum for the other two. The IndexArtificial Analysis’s Intelligence Index combines nine evaluations the firm runs itself, among them Terminal-Bench 4.0 and HLE-Diamond.

Guardrails

Alibaba wrote that it had sorted the model’s training data by purpose. “For cybersecurity, we filter data according to the specific purpose of each task, retaining data for defensive tasks such as vulnerability verification, detection, and remediation, while removing data for offensive tasks aimed at enhancing attack capabilities,” the announcement said. After a review by its own security experts, it said: “No capabilities posing large-scale cybersecurity risks have been observed.”

The company also showed the model at the laboratory bench. In a 48-hour run on Deep Principle’s Mira platform, Alibaba says, Qwen4 designed a coolant molecule for data centers and directed an automated lab to make it, “ultimately yielding a physical sample of the material.”

Alibaba was not alone. Tencent also released Hy4 Flash, a 224-billion-parameter model priced at 7 cents per million input tokens, the Post reported. Two of China’s largest technology companies thus put low-priced models on sale on the same day in China, in the week Anthropic plans to price its stock offering, on Wednesday, according to its European prospectus. What to watch next is the open Qwen4 weights, which Alibaba has promised but not dated.

Sources

  1. Qwen Team (Alibaba), “Qwen4-Max: Your New Teammate Has Arrived.”
  2. QwenCloud (Alibaba), “Qwen4-Max — model page and pricing.”
  3. QwenCloud (Alibaba), “API Pricing (Singapore region).”
  4. Artificial Analysis, “Qwen4 Max — Intelligence, Performance & Price Analysis.”
  5. South China Morning Post (owned by Alibaba), “Alibaba launches Qwen4-Max model with strict cybersecurity guardrails,” Nov. 16, 2026.
  6. Anthropic PBC (S.E.C. filing), “Prospectus for the offering in the European Economic Area (free writing prospectus), §8.1.13, expected timetable,” Nov. 12, 2026.