Alibaba Drops Qwen 3.6 Max Preview—Its Most Powerful Model Yet

Alibaba released a preview of its adjacent flagship AI exemplary connected Monday, pushing the Qwen bid further into the frontier race. Qwen 3.6-Max-Preview is the astir almighty exemplary the institution has shipped truthful far, topping six large coding benchmarks and posting meaningful gains successful satellite cognition and acquisition pursuing implicit its predecessor, Qwen 3.6-Plus.

The exemplary is disposable present done Qwen Studio and the Alibaba Cloud Model Studio API nether the drawstring qwen3.6-max-preview. It is simply a proprietary, hosted exemplary with nary unfastened weights, and its API is compatible with some OpenAI and Anthropic specifications, meaning developers tin plug it into existing pipelines with minimal changes.

This is besides a displacement successful Alibaba’s concern model, arsenic it was known to supply almighty open-source models by default. The little extremity models are inactive unfastened source.

According to Qwen’s official blog, Qwen 3.6-Max-Preview ranked archetypal crossed respective large benchmarks that trial coding and cause capabilities, including SWE-bench Pro (real-world bundle engineering tasks), Terminal-Bench 2.0 (command-line execution), SkillsBench (general problem-solving), QwenClawBench (tool use), QwenWebBench (web interaction), and SciCode (scientific programming).

In presumption of gains implicit Qwen 3.6-Plus, benchmarks for agentic skills enactment it connected apical of the household and different models similar Calude 4.5 oregon GLM 5.1. The exemplary besides improved successful cognition benchmarks, with SuperGPQA (advanced reasoning) expanding by 2.3% and QwenChineseBench (Chinese connection performance) by 5.3%. Instruction-following ability, measured by ToolcallFormatIFBench, enactment it connected apical of the rankings, beating Claude.

The merchandise lands 3 days aft Alibaba open-sourced Qwen 3.6-35B-A3B, a 35-billion-parameter exemplary that activates lone 3 cardinal parameters per inference.

That attack is designed to chopped compute costs without sacrificing output quality. Combined with Monday's Max release, the Qwen 3.6 lineup present spans Max-Preview astatine the apical of the family, the Qwen Plus variance for balanced workloads, Flash for speed-first tasks, and 35B-A3B for those moving locally.

The Max-Preview besides ships with preserve_thinking, a diagnostic that carries reasoning traces crossed multi-turn conversations. Alibaba specifically recommends it for agentic tasks wherever discourse continuity matters. For developers moving autonomous agents oregon long-running codification procreation workflows, that is simply a meaningful addition.

As Decrypt reported past week, Alibaba unopen down the escaped tier of Qwen Code conscionable days aft chap Chinese laboratory MiniMax rewrote its open-source licence to artifact commercialized usage without written authorization. Both moves awesome a broader shift: Chinese AI labs that built monolithic adoption connected free, unfastened services are present pivoting toward monetized, proprietary offerings. Qwen overtook Meta's Llama arsenic the astir deployed self-hosted exemplary connected the planet—and that momentum was built astir wholly connected escaped access.

That free-to-paid modulation runs parallel to different trend. Chinese unfastened models went from 1.2% of planetary open-model usage successful precocious 2024 to astir 30% by extremity of 2025, with Qwen starring the charge. The Qwen 3.6-Max-Preview is the proprietary extremity of that spear arsenic it’s the exemplary Alibaba is betting volition vie straight with OpenAI’s GPT and Anthropic’s Claude astatine the frontier.

Qwen 3.6-Max-Preview is explicitly labeled a enactment successful progress. Alibaba said the exemplary is inactive nether progressive improvement and expects further gains successful aboriginal versions. Independent benchmarking from Artificial Analysis puts it arsenic the 2nd champion performing exemplary down Muse Spark—well supra the median of comparable reasoning models successful its terms tier. The exemplary supports a 256k token discourse model and handles substance only, with nary representation input astatine launch.