Alibaba's Qwen 3.6-Max-Preview Tops Six Coding Benchmarks

Alibaba's Qwen team released Qwen 3.6-Max-Preview on April 20, a new closed-weight model that topped six major coding benchmarks at once.

The model finished first on SWE-bench Pro, Terminal-Bench 2.0, and SkillsBench, among others. It supports a 256,000-token context window, which means developers can load entire codebases, documentation, and test suites into a single session without chunking them across multiple calls.

Qwen 3.6-Max-Preview is hosted as a proprietary API product, a shift from Alibaba's earlier practice of releasing open weights across its full lineup. Smaller Qwen models remain open-source. Other labs have made the same calculation: keep the most capable models behind a paid API, make the lighter ones freely available.

That puts Alibaba in direct competition with Anthropic's Claude Opus 4.7, which recently retook the coding benchmark lead, and OpenAI's GPT-5.4 family. Topping six benchmarks at once is a strong showing, though benchmark scores and real-world coding performance don't always line up cleanly.

Qwen 3.6-Max-Preview is available through Alibaba Cloud's inference API. Pricing details have not been announced.