Alibaba previews Qwen3.8, a 2.4T model
Qwen3.8-Max-Preview lands on Alibaba's Token Plan and Qoder two days after Kimi K3, claiming second place behind Claude Fable 5 on the team's own evals.
Alibaba’s Qwen team announced Qwen3.8 on July 19, posting that it’s “launching and going open-weight soon” without committing to a release date. A preview build, Qwen3.8-Max-Preview, is live now on Alibaba’s Token Plan, Qoder, and QoderWork. Qwen developer Shuai Bai described it as the team’s first multimodal model above one trillion parameters, at 2.4 trillion total, handling text, images, video, and documents in one model.
Alibaba says Qwen3.8 should beat its predecessor at coding, full-stack development, data analysis, and office-style workflows, and that its internal evaluations place it just behind Claude Fable 5 and ahead of everything else on the board. None of that is independently verified. There’s no third-party benchmark run yet, and Alibaba’s own numbers are the only numbers in circulation.
The timing tells its own story: this preview landed two days after Moonshot’s Kimi K3, a 2.8-trillion-parameter model, went live API-only during the same World AI Conference window in Shanghai. Chinese labs are now racing each other’s release cadence as much as they’re racing the US frontier labs, and Alibaba is explicit that open weights for Qwen3.8 are still to come, the same gap K3 currently has. If you’re evaluating open models for a coding or agent workload, wait for weights and an independent eval before moving anything off GLM-5.2 or your current baseline on the strength of a vendor preview post.