Alibaba just rolled out Qwen3.8-Max, the biggest model it has ever built: 2.4 trillion parameters total, though only 95 billion switch on for any given request. That's the mixture-of-experts trick — instead of firing every parameter for every query, the system wakes only the specialist blocks it needs, which keeps compute costs down even as the model balloons in size.
Qwen3.8-Max reads text, images and video, and holds up to a million tokens in context — roughly 750,000 words, enough to load an entire book and ask it to find one paragraph. Alibaba built it for autonomous coding and long research runs, tasks where the model has to keep working on its own for dozens of steps without a human checking in.
On the Arena.AI leaderboard, it immediately topped every Chinese model on text tasks and landed second worldwide on vision, trailing only a variant of Anthropic's Claude Fable 5. Alibaba's own line is blunt: the model is "second only to Fable 5." That's the company's claim, not an independent finding, and on other benchmarks Qwen3.8-Max looks more like it's trading blows with Moonshot AI's Kimi K3 than clearly beating it.
Alibaba says the open-weights release lands next week, marking a return to giving away its top models after keeping recent flagships proprietary. For now it's live through Model Studio's API and the new QwenWork platform, which entered public beta the same day. Alibaba's Hong Kong-listed shares jumped 7% on the news, closing at HK$125.20.
The bigger picture is a race among Chinese labs to close the gap with US leaders by giving developers free open weights while Anthropic, OpenAI and Google keep their best models locked down.
