On Monday, Alibaba introduced the preview of its latest flagship AI model, marking a significant advancement in the Qwen series. The new Qwen 3.6-Max-Preview stands as the company’s most potent model yet, surpassing six major coding benchmarks and demonstrating notable improvements in global knowledge and instruction adherence compared to its predecessor, Qwen 3.6-Plus.
Alibaba announced on Twitter: 🚀 “Introducing Qwen3.6-Max-Preview, an early preview of our next flagship model Highlights：⚡️ Enhanced agentic coding ability over Qwen3.6-Plus📖 Robust world knowledge and instruction following🌍 Improved real-world agent and knowledge reliability performance Smarter…”
— Qwen (@Alibaba_Qwen) April 20, 2026
The model is accessible via Qwen Studio and the Alibaba Cloud Model Studio API under the identifier qwen3.6-max-preview. As a proprietary hosted model with no open weights, its API supports both OpenAI and Anthropic specifications, allowing developers to integrate it into existing workflows with minimal adjustments.
This release signifies a shift in Alibaba’s business strategy, moving away from providing powerful open-source models by default. Nevertheless, the lower-tier models remain open source.
According to Qwen’s official blog, Qwen 3.6-Max-Preview leads several major benchmarks that evaluate coding and agent capabilities, including SWE-bench Pro (real-world software engineering tasks), Terminal-Bench 2.0 (command-line execution), SkillsBench (general problem-solving), QwenClawBench (tool use), QwenWebBench (web interaction), and SciCode (scientific programming).
Compared to Qwen 3.6-Plus, the new model excels in agentic skills benchmarks, placing it above its predecessors and rivals like Calude 4.5 or GLM 5.1. It also shows enhanced knowledge capabilities, with a 2.3% increase in SuperGPQA (advanced reasoning) and a 5.3% boost in QwenChineseBench (Chinese language performance). In terms of instruction-following ability, measured by ToolcallFormatIFBench, it surpasses Claude.
The release follows closely on the heels of Alibaba open-sourcing Qwen 3.6-35B-A3B, a 35-billion-parameter model that activates only 3 billion parameters per inference. Parameters are crucial for an AI’s ability to learn, reason, and store information, with more parameters enabling broader knowledge.
This strategy aims to reduce computational costs without compromising output quality. Combined with the Max release on Monday, the Qwen 3.6 lineup now includes Max-Preview at the top, Qwen Plus for balanced tasks, Flash for speed-first tasks, and 35B-A3B for local use.
The Max-Preview also introduces preserve_thinking, a feature that maintains reasoning traces across multi-turn conversations. Alibaba suggests this feature for agentic tasks requiring context continuity. For developers managing autonomous agents or long-running code generation workflows, this addition is significant.
As Decrypt reported last week, Alibaba discontinued the free tier of Qwen Code shortly after fellow Chinese lab MiniMax revised its open-source license to restrict commercial use without permission. These actions reflect a broader trend: Chinese AI labs that once relied on free, open services are now transitioning toward monetized, proprietary offerings. Qwen surpassed Meta’s Llama as the most widely deployed self-hosted model globally—primarily due to free access.
This shift from free to paid models aligns with another trend: Chinese open models increased from 1.2% of global open-model usage in late 2024 to about 30% by the end of 2025, with Qwen leading the charge. The Qwen 3.6-Max-Preview represents the proprietary forefront as Alibaba’s contender against OpenAI’s GPT and Anthropic’s Claude at the frontier.
Explicitly labeled a work in progress, Alibaba notes that the model is still under active development and anticipates further enhancements in future versions. Independent benchmarking from Artificial Analysis ranks it second best behind Muse Spark, well above the median of comparable reasoning models in its price category. The model supports a 256k token context window and handles text only, with no image input at launch.