Alibaba launches Qwen4-Max, a 2.4tn-parameter AI model with closed weights

Alibaba says its new flagship cuts the average cost of a task by 58% in its own tests, but the weights stay closed. Smaller open Qwen4 models are promised.

Join the conversation
The illuminated Alibaba logo outside an office building in Beijing
The Alibaba logo outside an office building in Beijing. File photo. — TINGSHU WANG/REUTERS
Listen to this articleAlibaba launches Qwen4-Max, a 2.4tn-parameter AI model with closed weights3 min · device voice0:00 / 2:47

— Alibaba has released Qwen4-Max, the first flagship model built on its new Qwen4 architecture, and says it cuts the cost of completing a task by 58% in the company’s own tests. The model went live on Monday, 16 November.

Alibaba describes Qwen4-Max as a mixture-of-experts model with 2.4 trillion parameters in total, of which 71 billion are used for each token, plus 328 billion extra “N-gram embedding” parameters. It handles images natively and has a context window of one million tokens. The previous flagship, Qwen3.8-Max, used 95 billion active parameters.

In Alibaba’s Work360 evaluation, the average cost of a single task fell from 26.58 yuan to 11.16 yuan, and the overall score rose from 47.74 to 52.69. The company says token use dropped by about 24% to 35%, depending on the reasoning level. These are the company’s own figures and have not been independently verified.

Through its QwenCloud service, Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6 for the last flagship. Reused input served from a cache costs $0.075 per million tokens.

On public leaderboards, Qwen4-Max is the highest-ranked Chinese model on Arena’s agent table, placing 14th among individual models in a preliminary ranking, ahead of entries from Tencent, Moonshot AI, and DeepSeek. It debuted sixth globally on Arena’s Code Arena web-development ranking with 1,750 points. US labs still hold the top spots on both tables.

The catch is that the weights are closed. Developers can use Qwen4-Max through an API or Alibaba’s apps, but cannot download or run it themselves. Alibaba opened the weights of Qwen3.8-Max earlier this year, and says smaller Qwen4 models will be released with open weights, without giving sizes or dates.

Tencent’s Hunyuan unit launched a lightweight model, Hy4 flash, on the same day, with 224 billion parameters of which 12 billion are active. Alibaba and Tencent are due to report third-quarter results on 18 and 19 November. Qwen models are also due to power Apple Intelligence on iPhones in China from December, as The Thursday Times reported earlier this month.

TT News Desk

News from The Thursday Times' news desk.

More from this desk

The conversation

Comments are open on this story. An editor reads every one before it appears.

No one has said anything yet. Be the first.

Published by Thursday Media International thursday.media