Introducing Qwen3.6-Max-Preview: Enhanced World Knowledge and Coding Capabilities

I'm excited to share an early glimpse of our latest proprietary model, Qwen3.6-Max-Preview, which builds on the strengths of Qwen3.6-Plus with notable enhancements in world knowledge, instruction following, and agentic coding. This preview, available via Alibaba Cloud Model Studio, showcases improved performance across a variety of benchmarks, including SkillsBench, SciCode, and Terminal-Bench 2.0, with impressive gains in coding capabilities and knowledge reliability.

Qwen3.6-Max-Preview is designed for interactive use on Qwen Studio or through the Alibaba Cloud Model Studio API. One of its standout features is the `preserve_thinking` capability, which retains the reasoning content from previous interactions, making it particularly effective for agentic tasks.

Our evaluations reveal that Qwen3.6-Max-Preview outperforms its predecessor in several areas, achieving top scores on major coding benchmarks and demonstrating superior world knowledge and instruction adherence. For instance, it scores significantly higher on SuperGPQA and QwenChineseBench, indicating its enhanced understanding and application of complex information.

The model is still evolving, and we are committed to continuous improvement, anticipating further advancements in future iterations. We invite feedback from the community and are eager to see the innovative applications you develop with Qwen3.6-Max-Preview.

For those interested in integrating this model into their projects, the API supports industry-standard protocols and is compatible with OpenAI and Anthropic specifications. Detailed instructions and example code are available to facilitate seamless implementation.

Stay tuned for more updates as we refine Qwen3.6-Max-Preview, and don't hesitate to reach out with your insights and experiences.