SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Qwen3-Max-Thinking - Alibaba's flagship Qwen reasoning model

Author: AI小集 · Original source: AI工具集

What is Qwen3-Max-Thinking

Qwen3-Max-Thinking is Alibaba's newly launched flagship Qwen reasoning model, with over 1 trillion parameters and 36T Tokens of pre-training data. Through test-time scaling mechanisms and reinforcement learning, the model significantly improves reasoning performance and efficiency. Across 19 large model benchmark tests, the model has caught up with or even surpassedGPT-5.2-Thinking、Claude-Opus-4.5andGemini 3 Proand other TOP closed-source models. The model greatly enhances native Agent capabilities, can autonomously call tools, and provides smarter and smoother responses. Developers and ordinary users can useQwen Chat、QwenPC and web versions for free experience, and enterprises can obtain API (qwen3-max-2026-01-23) services throughAlibaba Cloud Model Studio. The Qwen App will also integrate the new model.

Qwen3-Max-Thinking

Main features of Qwen3-Max-Thinking

  • Powerful reasoning ability: It performs outstandingly in complex reasoning, mathematical computation, and logical analysis, and can handle highly difficult reasoning tasks, such as math competition problems and complex logic problems.
  • Autonomous tool calling: It has strong Agent capabilities and can autonomously call tools such as search engines and code interpreters to provide more accurate and smarter responses.
  • Efficient reasoning mechanism: Through Test-time Scaling technology, it improves reasoning efficiency, reduces redundant computation, and achieves more efficient reasoning results.
  • Multi-domain applications: It is suitable for multiple scenarios, including code generation, data analysis, content creation, etc., and can provide strong technical support for developers and users.

Technical principles of Qwen3-Max-Thinking

  • Ultra-large-scale parameters and data: The model has more than 1 trillion parameters, and the pre-training data volume is as high as 36T Tokens. Training with massive data improves the model's knowledge reserve and reasoning ability.
  • Test-time scaling mechanism: During the inference stage, it adopts an experience-accumulation-based, multi-round iterative scaling strategy to avoid redundant reasoning, focus on unresolved uncertainty, and significantly improve reasoning efficiency and performance.
  • Adaptive tool calling: Through a specially designed training process, the model can autonomously select and call built-in tools (such as search, code interpreters, etc.) according to task requirements, achieving smarter interaction and problem solving.
  • Reinforcement learning optimization: It conducts joint reinforcement learning training based on rules and model rewards on diverse tasks, further improving the model's performance and ability to align with human preferences.
  • Efficient reasoning architecture: By optimizing computational resource allocation and reasoning paths, it reduces computational redundancy, improves overall reasoning efficiency, and enables outstanding performance in complex tasks.

Performance of Qwen3-Max-Thinking

  • Handling highly difficult tasks: In multiple authoritative benchmark tests, Qwen3-Max-Thinking set a new global state-of-the-art (SOTA) record. For example, in the tool-enabled "Humanity's Last Exam" (HLE), Qwen3-Max-Thinking scored 58.3, far exceeding GPT-5.2-Thinking (45.5) and Gemini 3 Pro (45.8), becoming the highest score among all current models.
  • Mathematical reasoning: On IMO-AnswerBench, a mathematics ability test benchmark at the IMO difficulty level, Qwen3-Max-Thinking achieved the highest score of 91.5, demonstrating powerful mathematical reasoning ability.
Qwen3-Max-Thinking

How to use Qwen3-Max-Thinking

  • Developers canQwen ChatTry Qwen3-Max-Thinking for free
  • Enterprises can obtain API services throughAlibaba Cloud Model Studioto access API services
  • Regular users can try it onQwenPC and web versions. The Qwen app will soon be integrated, and all users can try it for free.

Application scenarios of Qwen3-Max-Thinking

  • Solving complex problems: Qwen3-Max-Thinking can efficiently solve highly difficult math, logic, and science problems, providing accurate reasoning and solutions.
  • Code generation and optimization: The model supports quickly generating high-quality code, optimizing programming workflows, and improving development efficiency.
  • Data analysis and report generation: The model can generate professional reports, provide data support for decision-making, and help enterprises gain insight into market trends.
  • Content creation: The model can help users write articles, creative copy, and stories, improving content creation efficiency.
  • Educational tutoring: Provides students with answers to academic questions and personalized learning suggestions, assists teaching, and improves learning outcomes.