Alibaba Releases Qwen3.8-Max AI Model With 2.4 Trillion Parameters

Following the announcement, Alibaba’s New York-listed shares rose 4.5% on Monday, while its shares on the Hong Kong exchange climbed 7%.

Technical Specifications and Performance

The Qwen3.8-Max is a multimodal model utilizing a sparse mixture-of-experts architecture. To maintain efficiency and reduce response times and computing costs, the system activates only 95 billion parameters per operation rather than the full set. The model supports a context window of up to 1 million tokens, enabling it to process thousands of pages of information. According to CNBC, the model was trained on 36T tokens.

View this post on Instagram about alibaba releases qwen3 model, Alibaba Qwen model

Alibaba shared benchmark results claiming that Qwen3.8-Max delivers comparable or occasionally better scores than Anthropic’s Fable 5. In the Vision Arena, the model ranks second only to Fable 5, while it ranks fifth in the Text Arena. While the company claims the model outperforms leading U.S. systems on several engineering, multimodal, and coding benchmarks, it trails on some general-purpose reasoning tests.

Comparisons with domestic rivals show Qwen3.8-Max ranking higher on several benchmarks than Moonshot AI’s Kimi K3, which was unveiled earlier this month. However, Kimi K3 remains China’s largest AI model with 2.8 trillion parameters.

Autonomous Capabilities and Use Cases

Alibaba is positioning the model for long-horizon autonomous work and visual intelligence. The company stated that Qwen3.8-Max can code highly autonomously with minimal human input. In one internal test, the model spent 16 days independently building and improving an AI coding tool by writing code, testing it, and fixing errors.

Other reported capabilities include:

• Professional Workloads: Conducting financial research, reviewing legal documents, and architectural 3D modeling.
• Visual Intelligence: Converting 100-hour livestreams, television series, or hundred-page documents into interactive, searchable knowledge hubs.

Market Strategy and Open-Source Release

In a strategic shift, Alibaba will release the Qwen3.8-Max weights for public download next week. This makes it the first Max-class Qwen model to be open-sourced, allowing users to customize the technology after the company had kept several recent flagship releases proprietary.

Alibaba has also adopted aggressive pricing to attract users. The model is priced at US$2 per one million input tokens and US$6 per million output tokens. While Freemalaysiatoday notes that DeepSeek’s latest release is the most affordable among new marquee models, Alibaba’s pricing is considered attractive compared to U.S. leaders.

Industry Context and Competition

The release follows a period of rapid activity among Chinese labs. Moonshot AI unveiled Kimi K3 on July 17, while ByteDance and MiniMax Group introduced new video generators on July 31. DeepSeek also recently expanded access to its V4 Flash model.

Vey-Sern Ling, managing director at Union Bancaire Privée, stated that the gap between Chinese and American AI models is narrowing fast, suggesting that investors often underestimate Chinese models due to general skepticism or U.S. chip restrictions. Wei Sun, a principal AI analyst at Counterpoint Research, estimates the general gap between Chinese and American models has narrowed to between three and six months, though this varies by task.

Despite the company’s claims, Nikkei Asia reported that some benchmark performance results indicate the model falls short of Alibaba’s previous claims.