Qwen2.5-Math-1.5B|7B|72B logo

Qwen2.5-Math-1.5B|7B|72B

Free

World's leading open-sourced mathematical LLMs

FreeFree tier
Type
Open Source
Company
Qwen Team

About Qwen2.5-Math-1.5B|7B|72B

Qwen2.5-Math is a series of open-sourced large language models specialized for mathematical reasoning, released by the Qwen Team on September 19, 2024. The series includes base models (1.5B, 7B, 72B parameters), instruction-tuned versions (Instruct), and a mathematical reward model (RM-72B). It supports both Chain-of-Thought (CoT) and Tool-integrated Reasoning (TIR) to solve math problems in English and Chinese. Trained on the Qwen Math Corpus v2 (over 1 trillion tokens), Qwen2.5-Math achieves significant performance gains over its predecessor Qwen2-Math on benchmarks such as GSM8K, MATH, and Chinese exam datasets (CMATH, Gaokao Math). The models are available under open-source licenses on GitHub, Hugging Face, and ModelScope.

Key Features

Supports Chain-of-Thought (CoT) and Tool-integrated Reasoning (TIR) for math problem solving
Available in three sizes: 1.5B, 7B, and 72B parameters
Instruction-tuned models (Instruct) and a dedicated reward model (RM-72B) included
Bilingual support for English and Chinese mathematics
Trained on Qwen Math Corpus v2 with over 1 trillion tokens
Significant performance improvements over Qwen2-Math on GSM8K, MATH, MMLU-STEM, CMATH, and Gaokao Math benchmarks
Open-sourced on GitHub, Hugging Face, and ModelScope

Pros & Cons

Pros
  • Open-source with multiple model sizes suitable for different resource constraints
  • State-of-the-art performance on major math benchmarks among open models
  • Supports both CoT and TIR for flexible reasoning strategies
  • Bilingual (English and Chinese) math problem solving
  • Includes a dedicated reward model for improved instruction tuning
Cons
  • Primarily designed for math problems; not recommended for general-purpose tasks
  • Larger models (72B) require significant computational resources
  • Limited to English and Chinese; no support for other languages in math reasoning
  • May still struggle with complex multi-step problems without TIR

Best For

Solving English and Chinese mathematical problems via CoT or TIREnhancing computational accuracy and symbolic manipulation in math tasksReinforcement learning and reward modeling for math reasoningEducational tools and automated math tutoring systemsResearch in mathematical reasoning and LLM alignment

FAQ

What languages does Qwen2.5-Math support for math problems?
Qwen2.5-Math supports solving math problems in English and Chinese.
What reasoning methods does Qwen2.5-Math use?
It supports Chain-of-Thought (CoT) and Tool-integrated Reasoning (TIR) for math problem solving.
Is Qwen2.5-Math open source?
Yes, the model weights are open-sourced and available on GitHub, Hugging Face, and ModelScope.
What model sizes are available?
Three sizes: 1.5 billion, 7 billion, and 72 billion parameters, each with base and instruction-tuned variants.
Can Qwen2.5-Math be used for non-math tasks?
The developers recommend using this series only for solving English and Chinese math problems; it is not designed for other tasks.