Alibaba's Qwen previews Qwen3.6-Max model

Alibaba's Qwen team released an early preview of Qwen3.6-Max, its next proprietary model. The release follows Qwen3.6-Plus, and the team says the model is still under active development, with further gains expected in later versions.

Compared with Qwen3.6-Plus, Qwen reports the biggest improvements in agentic coding: SkillsBench up 9.9 points, SciCode up 6.3, NL2Repo up 5.0 and Terminal-Bench 2.0 up 3.8. World-knowledge scores also rose, with SuperGPQA up 2.3 and the in-house QwenChineseBench up 5.3. On instruction following, ToolcallFormatIFBench went up 2.8.

The team says the preview scores highest on six coding benchmarks: SWE-bench Pro, Terminal-Bench 2.0, SkillsBench, QwenClawBench, QwenWebBench and SciCode. Two of those, QwenClawBench and QwenWebBench, are Qwen's own benchmarks. The post compares the model against "leading frontier models" but doesn't list those models or give absolute scores. The results are self-reported and haven't been independently verified.

Developers can call the model as qwen3.6-max-preview through the Alibaba Cloud Model Studio API, which offers OpenAI-compatible chat completions and responses endpoints as well as an Anthropic-compatible interface. It can also be tried on Qwen Studio. The model has a preserve_thinking option that keeps reasoning content from all earlier turns in the conversation, which Qwen recommends for agentic tasks.

Qwen didn't say anything about pricing, context length, model size or open weights, and it gave no date for a full release.