Alibaba releases Qwen3.8-Max-0902: #1 on Code Arena WebDev, 2.4T parameters

Alibaba released Qwen3.8-Max-0902 on September 2, a post-training refresh of Qwen3.8-Max. The architecture matches the August 3 release: 2.4T total parameters in a MoE design with roughly 95B activated parameters, a 1M-token context window, and native text, image, and video input. Pricing is unchanged at $2 input and $6 output per 1M tokens, with cache hits at $0.17 explicit and $0.25 implicit per 1M tokens.

The 0902 build focuses on coding and collaborative work, covering complex engineering code, multi-tool agent collaboration, and visual understanding for chart reasoning and document parsing. On the same day, the Code Arena: WebDev leaderboard updated with Qwen3.8-Max-0902 debuting at #1 with 1691 points, three points ahead of Claude Opus 5 (Max) at 1688, 17 points ahead of Kimi K3 (Max), and 22 points ahead of the prior Qwen3.8-Max. The model also took #1 in the Data & Analytics and Consumer Product sub-boards, the first non-Anthropic model to top a vote-based coding leaderboard this year.

On TerminalBench, DeepSWE, and NL2Repo, 0902 still trails the strongest frontier models, but it leads on repository understanding, automation benchmarks, and two visual reasoning tests. Blended price is roughly $5 per 1M tokens, the best position on the leaderboard's Pareto frontier. Qwen3.8-Max remains available only via the QwenCloud API; the open-weights release promised for Qwen3.8-Max has not shipped as of early September.