The 2.4-trillion-parameter Qwen3.8-Max is now available via API, with open weights set to follow next week
SOPA Images / Getty Images
Alibaba released Qwen3.8-Max on Monday, describing it as the most capable model in its Qwen family to date, with 2.4 trillion total parameters and 95 billion active parameters. The model is available through QwenCloud, with open weights planned for release next week — marking the first time Alibaba has open-sourced a model at this scale.
Alibaba stock rose 4.5% in premarket trading in New York and 7% on the Hong Kong exchange following the announcement.
Each query can draw on a context window of up to 1 million tokens — enough to cover approximately 750,000 words of input, according to the South China Morning Post. Alibaba described its capabilities as spanning coding, real-world work tasks, long-horizon autonomous operation, and multimodal understanding of documents, video, and images.
On coding, Alibaba said the model ran for more than 16 days in a fully autonomous operation, accumulating 265 commits, 127 pull requests, and 151 issues in a self-built software repository without human input. In a separate test, it entered a competition against 526 human teams, completing tasks under a 24-hour limit and finishing ahead of 87% of the field.
For document and video work, the model can ingest financial reports exceeding 200 pages and footage running beyond 100 hours, then compile that material into structured knowledge bases that users can search, the company said. It can also rebuild software applications by reading screenshots and turn two-dimensional architectural floor plans into three-dimensional renderings.
Alibaba published benchmark comparisons showing Qwen3.8-Max performing at comparable levels to Anthropic's Fable 5 on several coding and general agent tasks, and scoring above it on some multimodal and document benchmarks. On PaperBench, Qwen3.8-Max posted a score of 93.0 against Fable 5's 88.8. On the general capability benchmark IFBench, it scored 82.8 against Fable 5's 63.5. Fable 5 led on several other measures, including SWE-bench Pro and most visual agent tasks.
Alibaba previewed the model at the World Artificial Intelligence Conference in Shanghai last month under the name Qwen3.8-Max-Preview, at the time without supporting benchmark data. Monday's release includes a full benchmark table and technical documentation. The company also launched QwenWork, an all-in-one productivity platform, in public beta on the same day, targeting competitors including Tencent's WorkBuddy and Moonshot AI's Kimi Work, according to the South China Morning Post.
Alibaba's broader AI push includes a commitment to spend more than 380 billion yuan ($53 billion) on cloud and AI infrastructure over three years.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
