Qwen3.6-Plus: Toward Real-World Agents

Qwen3.6-Plus is the next-generation hosted model following the Qwen3.5 release, and is immediately available via API on Alibaba Cloud Model Studio. It provides a default 1M context window, with the core focus being significant enhancements to agentic coding and multimodal reasoning.

In the coding agent domain, it shows strong performance on benchmarks closer to real development work. It recorded SWE-bench Verified 78.8, SWE-bench Multilingual 73.8, SWE-bench Pro 56.6, and Terminal-Bench 2.0 61.6, emphasizing improvements in complex terminal tasks and automated execution as well.

Performance gains are also notable in general agent and tool usage capabilities. It ranks near the top with TAU3-Bench 70.7, DeepPlanning 41.5, MCPMark 48.2, MCP-Atlas 74.1, and WideSearch 74.3, explaining that it has jointly improved long-horizon planning and tool-calling abilities.

Fundamental knowledge and reasoning capabilities have also been maintained or improved. With results such as MMLU-Pro 88.5, GPQA 90.4, LiveCodeBench v6 87.1, and AIME26 95.3, it emphasizes broad coverage including long-context understanding, STEM reasoning, and multilingual support.

On the multimodal side, it extends to document understanding, visual reasoning, video understanding, and visual agents. Representative results include OmniDocBench1.5 91.2, RealWorldQA 85.4, VideoMME 87.8, and OSWorld-Verified 62.5, highlighting more stable performance in real work-type scenarios rather than simple recognition.

On the API side, a new preserve_thinking option has been added. It can maintain thinking content from previous turns, improving consistency in multi-step agent tasks and reducing redundant reasoning; the default value is false.

• API compatibility: Supports OpenAI specification-based chat completions / responses API, and an Anthropic-compatible interface
• Tool integration: Can be integrated with OpenClaw, Claude Code, Qwen Code, Kilo Code, Cline, and OpenCode
• Use cases: Frontend development, 3D scenes, games, complex repo-level problem solving, and multimodal work automation