Breaking Through the Limits of Qwen3-Max-Thinking

Qwen3-Max-Thinking is Qwen's flagship reasoning model, which strengthens knowledge, complex reasoning, instruction following, alignment, and agent capabilities across the board through parameter scaling and large-scale RL. Across 19 benchmarks, it showed performance comparable to GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3 Pro, and on core reasoning metrics it surpassed Gemini 3 Pro in some areas.

Two additional improvements were made. Adaptive tool-use automatically leverages Search, Memory, and Code Interpreter as needed during a conversation, without the user having to select a tool. Search and Memory reduce hallucination and enable real-time information and personalized responses, while Code Interpreter assists with code execution and calculation-based reasoning.

On benchmarks, it recorded MMLU-Pro 85.7, GPQA 87.4, LiveCodeBench v6 85.9, SWE Verified 75.3, Arena-Hard v2 90.2, and Tau² Bench 82.1, among others. It also showed AA-LCR 68.7 on long-context tasks, underscoring that it raises the bar across knowledge, reasoning, agent capability, and tool use together.

To improve reasoning performance, test-time scaling applied experience-accumulating, multi-round strategies. Instead of simply increasing the number of parallel paths N, remaining compute is reallocated to iterative self-reflection, and a take-experience mechanism compresses only the key insights from previous rounds to carry forward into the next round. As a result, with similar token consumption, it achieved improvements of GPQA 90.3→92.8, HLE 34.1→36.5, LiveCodeBench v6 88.0→91.4, IMO-AnswerBench 89.5→91.5, and HLE (w/ tools) 55.8→58.3.

On the practical usage side, Qwen Chat and the API have been released. The API name is qwen3-max-2026-01-23, and it can be used in an OpenAI-compatible way by setting base_url, with thinking mode enabled via enable_thinking=True in extra_body.

It is also compatible with the Anthropic API protocol, allowing integration with Claude Code. By setting the ANTHROPIC_MODEL, ANTHROPIC_SMALL_FAST_MODEL, ANTHROPIC_BASE_URL, and ANTHROPIC_AUTH_TOKEN environment variables, Qwen3-Max-Thinking can be directly connected to coding workflows.