Qwen3-Max-Thinking Release Notes

Qwen has announced Qwen3-Max-Thinking, a flagship reasoning model designed for high-performance complex reasoning, factual knowledge, and agentic capabilities. By scaling model parameters and utilizing extensive reinforcement learning, the model achieves performance comparable to leading industry models including GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3 Pro across 19 established benchmarks.

Technical Innovations

Adaptive Tool-Use Capabilities

Qwen3-Max-Thinking implements adaptive tool-use, allowing the model to autonomously select and invoke built-in Search, Memory, and Code Interpreter tools without manual user selection. This capability was developed through a focused training process involving initial fine-tuning for tool use followed by training on diverse tasks using both rule-based and model-based feedback.

  • Search and Memory: These tools are used to mitigate hallucinations and provide access to real-time information and personalized responses.
  • Code Interpreter: This allows the model to execute code snippets and apply computational reasoning to solve complex problems.

Experience-Cumulative Test-Time Scaling

To improve reasoning performance during inference, Qwen introduced a multi-round test-time scaling strategy for the "heavy mode." Rather than increasing parallel trajectories ($N$), which often leads to redundant reasoning, the model uses a "take-experience" mechanism to guide iterative self-reflection.

This mechanism distills key insights from previous rounds, preventing the model from re-deriving known conclusions and focusing instead on unresolved uncertainties. This approach provides higher context efficiency than referencing raw trajectories.

According to Qwen, this strategy consistently outperforms standard parallel sampling and aggregation with similar token consumption. Observed performance gains include:

Benchmark Baseline $\rightarrow$ Scaled Performance
GPQA 90.3 $\rightarrow$ 92.8
HLE 34.1 $\rightarrow$ 36.5
LiveCodeBench v6 88.0 $\rightarrow$ 91.4
IMO-AnswerBench 89.5 $\rightarrow$ 91.5
HLE (w/ tools) 55.8 $\rightarrow$ 58.3

Performance Benchmarks

Qwen3-Max-Thinking demonstrates competitive performance across several key dimensions:

Knowledge and STEM

  • MMLU-Pro: 85.7
  • MMLU-Redux: 92.8
  • C-Eval: 93.7
  • GPQA: 87.4
  • HLE: 30.2

Reasoning and Agentic Coding

  • LiveCodeBench v6: 85.9
  • HMMT Feb 25: 98.0
  • HMMT Nov 25: 94.7
  • IMOAnswerBench: 83.9
  • SWE Verified: 75.3

Instruction Following and Alignment

  • Arena-Hard v2: 90.2 (evaluated by GPT-4.1)
  • IFBench: 70.9
  • MultiChallenge: 63.3

Tool Use and Planning

  • Tau² Bench: 82.1
  • BFCL-V4: 67.7
  • Deep Planning: 28.7
  • HLE (w/ tools): 49.8

Availability and Integration

Qwen3-Max-Thinking is available via Qwen Chat (chat.qwen.ai) and through the API (model name: qwen3-max-2026-01-23). The API is OpenAI-compatible and supports the enable_thinking parameter in the extra_body field to activate reasoning capabilities.

Additionally, the model is compatible with the Anthropic API protocol, allowing it to be used with Claude Code by configuring the ANTHROPIC_BASE_URL to https://dashscope.aliyuncs.com/apps/anthropic.

Sources