Alibaba open-sources Qwen3.8-Max, a 2.4T model

Alibaba released Qwen3.8-Max and plans to publish the model’s open weights on Hugging Face and ModelScope next week. The company described Qwen3.8-Max as ‘the most capable model it has ever built.’

The model contains 2.4 trillion total parameters, with about 95 billion parameters active at any one time through a switching mechanism intended to reduce compute needs. Parameters determine model behavior; the switching mechanism limits how many parameters run for each request to lower hardware requirements.

Alibaba posted setup instructions that allow Qwen3.8-Max to run inside rival coding tools, including Anthropic’s Claude Code and OpenAI’s Codex. The company also provided an API that supports both protocols.

On Alibaba’s internal benchmark table covering 31 text tests, Anthropic’s Fable 5 ranked first on 15 tests, OpenAI’s GPT-5.6 Sol ranked first on nine, and Qwen3.8-Max ranked first on seven. On 12 coding tests, Qwen3.8-Max won one. The model led most comparisons on multimodal tasks involving documents, video and spatial reasoning.

Alibaba provided a cost comparison for inference and reported that Qwen3.8-Max runs at roughly 30% of the expense of Anthropic’s Claude Fable 5 on comparable tasks.

The company shared several performance examples. Qwen3.8-Max spent 16 days autonomously building a coding tool, producing 265 commits, opening 127 pull requests and filing 151 issues without human typing. The model reproduced the results of a research paper in five days and exceeded the paper’s reported score by 2.7 points. In a 24-hour machine learning contest, Qwen3.8-Max finished ahead of 458 of 526 human teams.

The release marks a change in Alibaba’s recent product approach. In April the company removed the free tier for Qwen Code and shifted parts of the Qwen family toward paid, closed offerings. The public release of Max‑scale weights is the first time Alibaba has offered open weights for a Max‑class Qwen model.

Open-weight models from Chinese developers have risen in self-hosting share, moving from under 2% of tokens on OpenRouter in late 2024 to about 61% by mid-2026. Qwen has surpassed Meta’s Llama as the most self-hosted model by that measure.

Washington imposed export controls on some models, including Fable 5 and Mythos 5, in June. Chinese authorities have been reported to consider limits on exporting domestic models.

Tutorials and instructions for running Qwen3.8-Max inside Claude Code and Codex will accompany the weight files when they appear on Hugging Face and ModelScope next week.