Axi · Сквозной — фон контента (skin) · 16 кат.Axi · Сквозной — фон контента (skin) · 16 кат.

Alibaba Introduces AI Model Qwen4-Max with 2.4 Trillion Parameters

• Evgenia Sliv (edited: )
Alibaba Introduces AI Model Qwen4-Max with 2.4 Trillion Parameters

Alibaba has introduced Qwen4-Max, a flagship AI model with 2.4 trillion parameters and a context of up to 1 million tokens. According to the company, the new architecture allows for complex tasks to be performed with less computational resources. However, the presented efficiency metrics are primarily based on Alibaba's own testing.

What is known about the new Qwen4-Max model

The launch of Qwen4-Max was announced by the company on November 16, 2026. The model is built on the Mixture of Experts (MoE) architecture, which utilizes only a portion of the parameters when processing each token.

With a total of 2.4 trillion parameters, Qwen4-Max activates about 71 billion. According to Alibaba, this is 25% less than the previous model Qwen3.8-Max, with the same total number of parameters.

Qwen4-Max supports a context window of up to 1 million tokens, image and video analysis, as well as interaction with external tools. Key application scenarios include programming, document processing, interface creation, and executing long multi-step tasks.

The model is already available through QwenCloud and API. Alibaba also announced its integration into Qwen Studio, QwenWork, and Qoder.

Alibaba announced a reduction in AI task costs

According to the company's internal tests, Qwen4-Max outperforms Qwen3.8-Max in several metrics, including programming, AI agent operations, and visual information analysis.

In its own Work360 test, covering 721 specialized professional fields, the new model used 35% fewer tokens in High reasoning mode.

The average cost of completing a task in the provided calculation decreased from 26.58 to 11.16 yuan, approximately 58%.

Alibaba attributes the improvements to the updated MoE architecture and the reduction in activated parameters. The company also lowered the announced API rates: the cost of 1 million input tokens is $1.5, output tokens $4.5, and cached input tokens $0.075.

However, the results of internal tests should not be considered independent confirmation of Qwen4-Max's advantages in all usage scenarios.

Qwen4-Max AI agents tested in scientific research

Alibaba presented several examples of the new model's application in extended research tasks.

According to the company, in one experiment, an AI agent worked for 48 hours, performing sequential stages of development and experimental verification of a cooling fluid molecule for chips.

In another research scenario, the agent made 454 tool calls and 3239 evaluations to develop and optimize 24 candidate molecules for Alzheimer's disease research.

These results describe the model's capabilities in the experiments presented by Alibaba, but do not imply that the developed compounds have undergone full clinical testing.

Alibaba prepares open Qwen4 models

The company also announced plans to release more compact models of the Qwen4 family with open weights. The publication timeline and full list of such models have not yet been disclosed.

Alibaba paid special attention to security. Developers reported changes in training and additional model adjustments to reduce risks in cybersecurity, as well as the dangerous use of AI in chemical, biological, radiological, and nuclear fields.

Detailed independent assessments of the effectiveness of these protective mechanisms are not included in the presented announcement.

This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.

Popular news