Alibaba launches 2.4-trillion-parameter Qwen3.8-Max with a 1-million-token context window

Alibaba’s Qwen team launched Qwen3.8-Max on August 3, 2026, a 2.4-trillion-parameter mixture-of-experts model that activates about 95 billion parameters for any single query and supports a 1-million-token context window.

The release is the latest sign that China’s leading AI labs are shipping frontier-scale models on a fast cadence. Qwen3.8-Max is multimodal, accepting text, images, video and documents, and is available now to global developers through Alibaba Cloud Model Studio, with open weights scheduled to follow the next week.

Pricing runs $2 per million input tokens and $6 per million output tokens, dropping to $0.25 per million for cached input; during a preview period Alibaba is charging 10% of standard rates under its Token Plan. The 1-million-token window, which the company measures at up to 991,000 tokens of input, is aimed at long-document and codebase work that shorter-context models handle poorly.

Alibaba says Qwen3.8-Max ranked fifth in Text Arena and second in Vision Arena on the Arena.AI leaderboard, and became the top-ranked Chinese model for text tasks. Those standings are the company’s own account and reflect a crowdsourced preference benchmark rather than an independent capability test, so they say more about user-rated responses than about accuracy or safety.

Investors reacted quickly: Alibaba shares rose about 4% in New York premarket trading and roughly 7% in Hong Kong after the announcement. The open-weights release, due within a week, will be the real test of whether outside developers can reproduce the leaderboard showing on their own hardware.