Overview
- Alibaba is offering the preview as a proprietary hosted model that developers can call through Qwen Studio or Alibaba Cloud’s Model Studio API under qwen3.6-max-preview with endpoints that match OpenAI and Anthropic formats.
- Alibaba says the model leads major coding and agent tests such as SWE-bench Pro, Terminal-Bench 2.0, SkillsBench, QwenClawBench, and QwenWebBench, while independent firm Artificial Analysis ranks it second behind Muse Spark.
- The preview adds preserve_thinking to carry reasoning traces across multi-turn chats, supports a 256k-token context window, and launches with text-only input.
- Reported gains over Qwen 3.6-Plus include a 2.3% rise on SuperGPQA and a 5.3% jump on QwenChineseBench, and it topped ToolcallFormatIFBench for instruction following over rivals such as Claude.
- Alibaba frames the release as work in progress and pairs it with a new lineup strategy that keeps lower tiers open, including the recently open-sourced Qwen 3.6-35B-A3B for local use, while it retires the free Qwen Code tier to emphasize paid access.