Alibaba releases Qwen3.8-Max, a 2.4T-parameter MoE model with 1M-token context

Alibaba unveiled Qwen3.8-Max, a 2.4-trillion-parameter Mixture-of-Experts model that activates 95 billion parameters per query, supports text, image, and video input, and offers a 1-million-token context window. It is generally available through QwenCloud and Alibaba Cloud Model Studio, priced at $2.00 per million input tokens and $6.00 output, with cached input as low as $0.25. Open-weights releases of Qwen3.8-Max and a smaller Qwen3.8-27B are scheduled for August 12, 2026.

The model is already making competitive noise: Alibaba's Qwen account announced it ranks #5 on the Artificial Analysis Intelligence Index and #1 on the Agentic Index — a strong showing for agentic workloads that helped push its Hacker News coverage to 464 points. Independent commentator Simon Willison flagged particular excitement not for the giant Max model but for the upcoming 'laptop-sized' Qwen 3.8 open-weights variants, underscoring how much of Qwen's mindshare comes from local-deployable weights.

Alibaba paired the model launch with a platform consolidation, merging QoderWork, MuleRun, and Wukong into a unified QwenWork enterprise agent platform now in public beta — signaling a full-stack push from raw model to deployed agents. Forbes framed the release as evidence China is closing the capability gap with U.S. labs.

Competitively, Qwen3.8-Max lands in the middle of an intense China price-and-capability war alongside DeepSeek's V4-Flash, pressuring Western labs on per-token economics precisely as OpenAI makes Luna free and DeepSeek warns of hikes. The 1M-token context and aggressive cached-input pricing target agentic and long-document workloads where sustained context matters. The key watch item is the August 12 open-weights drop: if the smaller variants deliver frontier-adjacent quality at laptop scale, they could meaningfully shift the local-LLM landscape.