Alibaba Launches Most Powerful AI Model, Qwen 3.8-Max

Alibaba unveiled Qwen 3.8-Max, the most powerful iteration in its Qwen large language model series, on Monday. This new model features 2.4 trillion parameters, though it actively uses 95 billion parameters at any given time. It will be accessible through Alibaba Cloud's Model Studio, with open weights scheduled for release next week.

“

Qwen 3.8-Max sits close to the size of domestic rival Moonshot AI's Kimi K3, which launched last month with 2.8 trillion parameters.

Qwen 3.8-Max rivals the scale of domestic competitor Moonshot AI's Kimi K3, which launched last month with 2.8 trillion parameters. Both models possess the capability to process text, images, and video, and can handle up to one million tokens concurrently. On the crowdsourced comparison platform Arena.AI, Qwen 3.8-Max immediately secured the top position among Chinese models for text-based tasks, although it still trails behind several offerings from Anthropic, including Claude Fable 5. For vision tasks, it ranked second globally, surpassed only by a variant of Fable 5.

The model utilizes a mixture-of-experts architecture, which efficiently distributes tasks among specialized segments rather than engaging the entire system for every query. Alibaba states this approach helps reduce costs and response times. In contrast, major AI developers like OpenAI, Anthropic, and Google do not publicly disclose the parameter counts for their proprietary models.

“

The model uses a mixture-of-experts architecture, which divides work among specialised segments rather than activating the entire system for every query.

Alibaba highlighted Qwen 3.8-Max's ability to execute complex, multi-day projects autonomously. In one demonstration, the model spent over ten days independently developing a self-evolving software harness from scratch. This process involved incorporating user feedback, running its own tests, and iterating through code, previews, and logs without any human intervention. In another test, it successfully recreated a machine learning research paper from the ground up, requiring approximately 125 hours of GPU training across 33 rounds, generating 7,600 lines of code, and proposing 18 innovative improvements that ultimately surpassed the original paper's methodology. When entered into a real online competition against 526 human teams, it outperformed 87 percent of the participants.

“

In one test, the model spent over ten days autonomously coding a self-evolving software harness from scratch, incorporating user feedback, running its own tests, and iterating through code, previews, and logs without human help.

The company also showcased the model's versatility across various professional domains. This included completing a full legal compliance review in under an hour, generating an interactive banking app prototype with no revisions needed, and creating a 26-dish restaurant menu complete with calorie and cost analysis in a single pass.

“

Entered into a real online contest with 526 human teams, it beat 87 per cent of the field.