Alibaba Refreshes Qwen3.8-Max with a 0902 Snapshot, Takes No.1 on Frontend Coding
Alibaba’s Qwen team released Qwen3.8-Max-0902 on September 2, 2026, an updated snapshot of its flagship Qwen3.8-Max model, announced via the official Qwen account on X and rolled out immediately through the QwenCloud API and Alibaba Cloud’s Model Studio.
Details
- Same architecture, deeper post-training: the model keeps its 2.4-trillion-parameter Mixture-of-Experts design and 1-million-token context window unchanged, but receives further post-training focused specifically on coding and “Cowork”-style collaborative office tasks
- Coding leaderboard result: Qwen says the model’s CodeArena WebDev (frontend coding) score rose 22 points to 1,691, putting it in first place on that leaderboard, ahead of Claude Opus 5 Max (1,687) and Kimi K3 Max (1,674)
- Broader coding gains: Qwen reports improvements across all eight of its internal coding benchmarks, with the largest jumps on TerminalBench 3.0 (11.3 to 29.0) and ProgramBench Almost Solved (10.5 to 28.0), both more than doubling
- Agent and vision improvements: Alibaba also describes stronger multi-tool orchestration for long-horizon autonomous development, and sharper vision understanding for chart reasoning, document parsing, and multimodal perception
- Unchanged pricing and specs: same $2/million input and $6/million output token pricing, 991K max input tokens, 131K max output tokens, and up to 262K reasoning tokens in thinking mode
- Availability: live immediately via the model ID
qwen3.8-max-0902(dated aliasqwen3.8-max-2026-09-02) on QwenCloud and Alibaba Cloud Model Studio, and already integrated into Qwen’s own apps including Qwen Office and Qoder
What happened next
The 0902 refresh arrives about three weeks after Qwen3.8-Flash-Next previewed the architecture planned for the next-generation Qwen4, suggesting Alibaba is iterating quickly on its existing Max line while that larger architectural transition is still in progress. No release date for Qwen4 itself has been announced.