Curiosity · AI Model

Qwen2.5-Math 72B

Qwen2.5-Math 72B, released alongside the broader Qwen2.5 family in 2024, is Alibaba's heaviest math-specialised open-weights model. It supports both chain-of-thought and tool-integrated reasoning (executing Python for numeric steps) and reported state-of-the-art open-weights scores on MATH, GSM8K, and competition-level benchmarks at launch.

Model specs

Vendor
Alibaba
Family
Qwen2.5
Released
2024-09
Context window
4,096 tokens
Modalities
text

Strengths

  • State-of-the-art open-weights math performance at launch
  • Supports both CoT and Python tool-integrated reasoning
  • Broad coverage from grade-school to competition-level problems

Limitations

  • Specialist — weaker on general conversation than Qwen2.5 Instruct
  • 72B footprint requires multi-GPU serving
  • License constraints on certain large-scale commercial deployments

Use cases

  • Open-weights Olympiad and competition math tutors
  • Tool-integrated reasoning research with Python
  • Math-heavy RAG pipelines over textbooks and papers
  • Classroom math teaching assistants

Benchmarks

BenchmarkScoreAs of
MATH≈85%2024-09
GSM8K≈95%+2024-09
AIME 2024 (TIR)strong open-weights score2024-09

Frequently asked questions

What is Qwen2.5-Math 72B?

Qwen2.5-Math 72B is Alibaba's open-weights 72-billion-parameter math specialist, fine-tuned from Qwen2.5 with additional training on math-specific corpora and tool-integrated reasoning.

What is tool-integrated reasoning?

Tool-integrated reasoning (TIR) lets the model interleave natural-language steps with executed Python snippets, improving accuracy on numeric and symbolic problems where pure chain-of-thought struggles.

How does Qwen2.5-Math compare to Mathstral?

Qwen2.5-Math 72B is much larger and significantly stronger on hard benchmarks. Mathstral 7B is cheaper to run and fine-tune — a practical choice for classroom-scale deployments.

Sources

  1. Alibaba Qwen — Qwen2.5-Math blog — accessed 2026-04-20
  2. Hugging Face — Qwen/Qwen2.5-Math-72B-Instruct — accessed 2026-04-20

Related

  • Qwen2.5-VL 72B — Qwen2.5-VL 72B is Alibaba's top-tier open-weights vision-language model — a 72B transformer with agentic UI grounding, long-video understanding, and precise document OCR.
  • Qwen2-Audio 7B — Qwen2-Audio 7B is Alibaba's open-weights audio-language model — a 7B transformer that accepts speech, music, and environmental sounds and responds in natural-language text.
  • Mathstral 7B — Mathstral 7B is Mistral AI's open-weights math specialist — a 7B Mistral fine-tune aligned with Project Numina to solve Olympiad-style problems with chain-of-thought.
  • All AI Models — Browse the complete ai models library
  • VSET — Programmes offered — B.Tech programs (CSE, AI&ML, AI&DS, Electronics, IIoT)