A flagship model positioned at the top of China's open-weight field
Alibaba unveiled Qwen3.8-Max, describing it as the largest and most capable artificial-intelligence model it has built, with 2.4 trillion parameters. The company placed it in direct comparison with domestic rival Moonshot AI's Kimi K3, which carries 2.8 trillion parameters and launched the previous month. On the crowdsourced Arena.AI platform, Qwen3.8-Max became the highest-ranked Chinese text model and ranked second globally on the visual-analysis leaderboard, trailing only an Anthropic Claude variant.
Multimodal design and a long context window for agent work
Qwen3.8-Max is a multimodal foundation model that handles text, images and video and accepts up to 1 million tokens per prompt, equivalent to roughly 750,000 words. Alibaba said the model can process lengthy documents, video series and live streams to build searchable knowledge bases, and demonstrated tasks such as recreating software applications from screenshots, generating interactive games and educational animations, and converting two-dimensional floor plans into 3D visualisations. The company also said the model completed a software-engineering project over a 16-day autonomous run, a claim that has not been independently tested.
Mixture-of-experts architecture to control compute costs
Qwen3.8-Max uses a mixture-of-experts design that divides work among specialised components, activating only about 95 billion parameters for any given request. Alibaba framed the approach as a way to keep a very large model cheap to run and quick to answer, a relevant consideration as Chinese developers emphasise cost efficiency alongside raw scale. The model is positioned against a competitive backdrop in which Moonshot paused new sign-ups for Kimi K3 to protect capacity, underscoring how quickly strong Chinese models attract users.
Distribution through Model Studio and a new workplace agent
The model is available to global developers through Alibaba Cloud's Model Studio application programming interfaces, and on the same day Alibaba opened public beta for QwenWork, an all-in-one workplace AI agent platform delivered through web and desktop apps. QwenWork targets rival offerings including Tencent's WorkBuddy, Moonshot AI's Kimi Work, Anthropic's Claude Cowork and ChatGPT Work, and is also integrated with the Qwen Office agent built around the new model. Alibaba plans to follow with an open-weights release of Qwen3.8-Max next week and a separate open-source release of Qwen3.8-27B, marking a return to open-sourcing top-tier models after keeping several recent flagship releases proprietary earlier in the year. Alibaba's Hong Kong-listed shares rose 7 percent following the release, closing at HK$125.20 the same day. Token pricing for Qwen3.8-Max was not disclosed in the reporting.
Openness as a differentiator from closed Western labs
Alibaba's choice to publish parameter counts and to release open weights sits in contrast with OpenAI, Anthropic and Google, which do not disclose parameter counts for their closed-source models. The open approach has become a recurring pitch from Chinese developers, who frame transparency about size and architecture as a way to attract developers and accelerate adoption of domestic models.
What remains to be seen
The open-weights release scheduled for next week and the impending Qwen3.8-27B open-source release will be the next measurable test of whether Qwen3.8-Max's Arena.AI rankings hold once external developers can run the model directly. No token pricing has been published, leaving cost competitiveness against Kimi K3 and Western flagship models undefined in the current reporting.
Share this article
