Alibaba has rolled out a preview of its most advanced AI model yet, stepping up its push into the top tier of global AI development.
Summary
- Alibaba launched Qwen 3.6 Max Preview, its most advanced AI model, with top rankings across multiple coding and agent benchmarks.
- The model is offered as a proprietary hosted system, marking a move away from the company’s earlier open access approach.
According to an X post from Alibaba’s Qwen team, the new model, Qwen 3.6-Max-Preview, has taken the lead across several key benchmarks, particularly in coding and agent-based tasks. Internal testing placed it ahead on SWE-bench Pro for real-world software work, Terminal-Bench 2.0 for command-line execution, and SkillsBench for general problem-solving, alongside strong results in tool use and web interaction benchmarks.
Performance gains extend beyond coding. SuperGPQA scores rose by 2.3%, pointing to stronger reasoning ability, while QwenChineseBench improved by 5.3%, underlining better performance in Chinese language tasks. Instruction-following capability also ranked at the top in ToolcallFormatIFBench, where the model outperformed competing systems, including Claude.
The release is now live through Qwen Studio and Alibaba Cloud’s Model Studio API under the identifier qwen3.6-max-preview. Developers can integrate it without major changes, as the API supports both OpenAI and Anthropic formats.
Alibaba’s latest move signals a noticeable change in direction. Earlier versions of Qwen built momentum through open-source access, helping the model family gain widespread adoption. Max-Preview, however, is a hosted proprietary system with no open weights.
Lower-tier models remain open source, but the flagship tier is now positioned as a paid, controlled offering. The shift comes just days after Alibaba open-sourced Qwen 3.6-35B-A3B, a model designed to run efficiently by activating only 3 billion of its 35 billion parameters during inference, cutting compute costs while maintaining output quality.
Combined, the Qwen 3.6 lineup now spans multiple use cases. Max-Preview sits at the top for high-end workloads, while the Plus variant targets balanced tasks, Flash focuses on speed, and 35B-A3B supports local deployments.
A new feature introduced with Max-Preview, called preserve_thinking, allows the model to carry reasoning traces across multiple interactions. Alibaba recommends it for agent-driven workflows where maintaining context across long sessions is important.
Alibaba described the release as an ongoing project, noting the model is still under active development and likely to improve in future updates. Qwen 3.6-Max-Preview currently supports a 256k token context window and is limited to text input, with no image capabilities at launch.
Industry transition toward monetization
Alibaba recently shut down the free tier of Qwen Code, while MiniMax updated its open-source license to restrict commercial use without approval. Both actions point to a gradual move away from free access models that initially drove adoption.
Qwen’s growth has been notable. The model family overtook Meta’s Llama as the most widely deployed self-hosted system, with much of that traction built during its open-access phase. At the same time, Chinese open models expanded their share of global usage from 1.2% in late 2024 to around 30% by the end of 2025.
Max-Preview now sits at the center of Alibaba’s effort to compete directly with leading frontier models from OpenAI and Anthropic.
Independent analysis from Artificial Analysis ranks the model as the second-best performer behind Muse Spark, placing it well above the average for reasoning models in its pricing category.
