Alibaba’s Qwen team has previewed Qwen3.8-Max, a 2.4-trillion-parameter model the company calls its most capable system yet. The Chinese tech giant unveiled the preview on July 19 at the World AI Conference in Shanghai, describing it as “second only to Fable 5” — a reference to Anthropic’s flagship Claude model — among frontier AI systems.
What’s new
Qwen3.8 is the first model in Alibaba’s Qwen family above one trillion parameters to go beyond text, processing images, video, and documents alongside written prompts. It builds on the sparse mixture-of-experts architecture Alibaba has used across recent Qwen releases, and the company says it should beat the prior Qwen3.7-Max at coding, full-stack development, data analysis, and office tasks.
The preview, called Qwen3.8-Max-Preview, is live now on Alibaba Cloud’s Token Plan subscription service and its Qoder and QoderWork coding platforms, priced at 10% of the eventual standard rate. Alibaba has not disclosed the model’s activated-parameter count, released a model card, or published benchmark scores, so its ranking against other large language models cannot yet be independently verified.
Part of a wider race
The debut lands two days after Moonshot AI opened weights on Kimi K3, a 2.8-trillion-parameter model that briefly rattled chip stocks, and follows Apple’s move to route its China-market AI features through Qwen. Chinese labs have increasingly clustered around multi-trillion-parameter releases this year, a scale previously reached mainly by a handful of US frontier labs.
Alibaba said an open-weight version of Qwen3.8 is planned but gave no release date, license, or standard API price. The company’s Hong Kong-listed shares rose following the announcement.