Alibaba ships Qwen3.8-Max: 2.4 trillion parameters, 95B active, open weights promised
Alibaba released Qwen3.8-Max on 3 August — a 2.4-trillion-parameter mixture-of-experts model activating 95 billion at a time, handling text, images and video across a million-token context. It went live on Model Studio with open weights said to follow. It is the largest model the Qwen family has produced.
This is not an incremental point release, and the version number undersells it badly. A 2.4-trillion-parameter MoE activating 95 billion per token is a flagship architecture bet, and on the crowdsourced Arena leaderboard it immediately became the highest-ranked Chinese model for text and second globally for vision.
The sparsity ratio is the engineering story. Activating roughly 4% of parameters per token is what makes a model this size servable at all — total capacity scales with the parameter count while inference cost tracks the active set. That is the same bet Mistral made with Large 3, executed at four times the scale.
The open-weight question is the one that matters for everyone downstream. It shipped API-first through Model Studio with weights promised to follow, which has become the standard sequence across every lab running both an open and a commercial line. Promised is not published — but Alibaba has generally delivered, which is why the promise is worth tracking rather than dismissing.
MarkTechPost — Alibaba Qwen releases Qwen3.8-Max: a 2.4 trillion parameter MoE model → · The Daily Star — Alibaba releases Qwen 3.8-Max, its largest AI model yet → · TestingCatalog — Alibaba released Qwen3.8-Max with open weights coming soon →