scmp.com web signal

Alibaba Opens Qwen3.8-Max to Global Users, Weights Next Week

TL;DR

  • Alibaba made Qwen3.8-Max, a 2.4-trillion-parameter multimodal model with a 1 million-token context window, widely accessible to global users.
  • Open weights for Qwen3.8-Max are scheduled to release the following week, reversing Alibaba's recent pattern of keeping flagship models proprietary.
  • QwenWork, an all-in-one workplace AI agent platform, entered public beta the same day, targeting Tencent's WorkBuddy, Moonshot AI's Kimi Work and Claude Cowork.

A frontier-scale Chinese model going open-weights is the part of this release worth reading past the headline for. Alibaba made Qwen3.8-Max, a 2.4-trillion-parameter multimodal model, widely accessible to global users this week and said the open-weights version will follow next week. The South China Morning Post reported that the model supports a context window of up to 1 million tokens and can take in lengthy documents, television series and live streams as inputs, with showcased tricks like recreating software applications from screenshots and converting two-dimensional floor plans into 3D visualisations.

Alibaba pushed a companion product into public beta the same day, an all-in-one workplace agent called QwenWork, distributed as web and desktop apps. According to the SCMP write-up, it lines up directly against Tencent's WorkBuddy, Moonshot AI's Kimi Work and imported rivals including Claude Cowork and ChatGPT Work. That is a crowded field, and one where distribution, price and enterprise integrations will probably decide it more than raw model quality.

The part that lifts this out of a routine release note is the open-weights pivot. Alibaba had recently kept several of its flagship releases proprietary; putting a frontier-scale model back into the open pool is a strategic reversal rather than a housekeeping choice. For teams outside China who have been paying for hosted access to models in this class, a downloadable checkpoint reshapes the build-versus-buy math for anyone with the compute to serve it.

The honest caveats are worth naming. The reporting does not include third-party benchmarks against Western frontier models, does not disclose the license the weights will ship under, and does not say what hardware you would need to actually serve a 2.4-trillion-parameter model at reasonable latency. The multimodal demos come from the company's own description, not an independent test, so take the specifics as reported, not settled.

If the weights land next week under a workable license, the interesting downstream is not the leaderboard fight, it is which Chinese enterprises and non-US developer teams start building on a frontier open model that Alibaba is now actively marketing to them.