DeepSeekMath

DeepSeekMath is DeepSeek AI’s open-source language model project focused on mathematical reasoning. Its headline release, DeepSeekMath 7B, showed that a relatively small open model — continually pre-trained on a large corpus of math-heavy web data and then aligned with reinforcement learning — could rival much larger proprietary systems on competition-level mathematics benchmarks.

Overview

DeepSeekMath is an open-source project on GitHub with about 3,500 stars, released by DeepSeek AI in 2024. The repository provides the DeepSeekMath-Base and instruction/RL-tuned 7B models, evaluation code and documentation of the training recipe. Notably, the project introduced Group Relative Policy Optimization (GRPO), a memory-efficient reinforcement learning algorithm that went on to power later DeepSeek models, including the reasoning-focused R series. The base model builds on DeepSeek-Coder-Base-v1.5 and 120B math-related tokens, achieving strong results on benchmarks such as MATH, where the RL model scored around 51.7% — approaching GPT-4 territory at a fraction of the size.

Key Features

• Open 7B models specialised for mathematical reasoning
• Introduces GRPO, an efficient RL algorithm for LLM alignment
• Documented data pipeline built from 120B math-focused tokens
• Strong benchmark results on MATH and other reasoning suites
• Evaluation code and model weights freely available
• Foundation for later DeepSeek reasoning models

💳 Pricing

DeepSeekMath is completely free and open source under the MIT license (models are released under DeepSeek’s model license). You can download the weights and run them on your own hardware at no cost.

Who Is It For

DeepSeekMath is for researchers and developers interested in mathematical reasoning with LLMs — whether you want a capable open math model, a case study in data curation, or the original GRPO implementation to learn from.