Qwen3.6-Max-Preview: Smarter, Sharper, and Still Evolving

Released after Qwen3.6-Plus, Qwen3.6-Max-Preview is an early preview of the next-generation proprietary model.

Compared to Qwen3.6-Plus, it has been improved in the following areas.

• agentic coding: SkillsBench +9.9, SciCode +6.3, NL2Repo +5.0, Terminal-Bench 2.0 +3.8
• world knowledge: SuperGPQA +2.3, QwenChineseBench +5.3
• instruction following: ToolcallFormatIFBench +2.8

In terms of availability, it is provided as a hosted proprietary model through Alibaba Cloud Model Studio, and can be called via API as qwen3.6-max-preview. It can also be tried interactively in Qwen Studio.

This release supports the preserve_thinking feature, which is recommended for retaining thinking content from previous turns in agentic tasks.

In the benchmark summary, it states that the model achieved the best scores on 6 coding benchmarks: SWE-bench Pro, Terminal-Bench 2.0, SkillsBench, QwenClawBench, QwenWebBench, SciCode.

However, it is explicitly stated that this model is still at the preview stage and is continuing to be improved. It explains that further enhancements are expected in the next version.