# DeepSeekMath

> The original 7B open math model that introduced GRPO — the RL method later used to train DeepSeek-R1.

- Maker: DeepSeek
- Line: DeepSeek Math
- Canonical page: https://ai-tldr.dev/models/deepseek-math/
- Updated: 2026-06-18

## In short

DeepSeekMath is a 7-billion-parameter open-weight math model from DeepSeek, first released in February 2024 (with the v3 paper revision following in April 2024). Rather than training from scratch, DeepSeek continued pre-training its DeepSeek-Coder-Base-v1.5 7B checkpoint on 120 billion math-related tokens scraped from Common Crawl, plus natural-language and code data.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 4K (4,096 tokens) | [1] |
| Licence | MIT (code) + DeepSeek Model License (weights); commercial use permitted · open weights | [1] |
| Released | April 2024 | [2] |
| Parameters | 7B (dense) | [1] |

## Sources

1. [deepseek-math-7b-rl model card (Hugging Face)](https://huggingface.co/deepseek-ai/deepseek-math-7b-rl)
2. [DeepSeek-Math GitHub repository](https://github.com/deepseek-ai/DeepSeek-Math)

## Overview

DeepSeekMath is a 7-billion-parameter open-weight math model from DeepSeek, first released in February 2024 (with the v3 paper revision following in April 2024). Rather than training from scratch, DeepSeek continued pre-training its DeepSeek-Coder-Base-v1.5 7B checkpoint on 120 billion math-related tokens scraped from Common Crawl, plus natural-language and code data. The result was a small model whose mathematical reasoning rivaled far larger systems: DeepSeekMath approached the MATH-benchmark level of Gemini-Ultra and GPT-4 while being open and self-hostable.

DeepSeek shipped three variants: DeepSeekMath-Base 7B (the continued-pretrained foundation), DeepSeekMath-Instruct 7B (chain-of-thought instruction-tuned), and DeepSeekMath-RL 7B (the strongest, refined with reinforcement learning). The headline result — 51.7% on the competition-level MATH benchmark and 88.2% on GSM8K using chain-of-thought without any external tools — came from the RL variant and beat every open-source model from 7B to 70B at the time. All variants run in a 4,096-token context.

DeepSeekMath's most lasting contribution is GRPO (Group Relative Policy Optimization), the reinforcement-learning algorithm introduced in this paper. GRPO is a memory-efficient variant of PPO that drops the separate value/critic model and instead estimates the baseline from a group of sampled outputs. It is the same method DeepSeek later scaled up to train DeepSeek-R1, making DeepSeekMath a direct technical ancestor of DeepSeek's reasoning models. The line continues today with the much larger DeepSeek-Math-V2.

## Benchmarks

- MATH (chain-of-thought, no tools) — DeepSeekMath-RL 7B: 51.7% ([source](https://arxiv.org/abs/2402.03300))
- MATH (self-consistency over 64 samples) — DeepSeekMath 7B: 60.9% ([source](https://arxiv.org/abs/2402.03300))
- GSM8K (chain-of-thought) — DeepSeekMath-RL 7B: 88.2% ([source](https://github.com/deepseek-ai/DeepSeek-Math))

## Strengths

- Introduced GRPO, the group-relative RL method DeepSeek later reused to train DeepSeek-R1
- Strong math reasoning from a small 7B model — 51.7% on MATH and 88.2% on GSM8K without external tools
- Continued pre-training on 120B math tokens from public web data, showing high-quality data engineering over raw scale
- Fully open weights with commercial-use license, in three variants (Base, Instruct, RL) for self-hosting and fine-tuning
- Approaches Gemini-Ultra / GPT-4-level MATH accuracy at a fraction of the parameter count

## Best for

- Self-hosted math problem solving and step-by-step (chain-of-thought) reasoning
- Research baseline for GRPO and reinforcement-learning-for-reasoning experiments
- Fine-tuning a small, efficient math foundation model for tutoring or STEM assistants
- Studying math-focused continued pre-training and data curation from web corpora
- Running competition-style math (GSM8K / MATH) evaluation on commodity hardware

## How to access

- [Hugging Face (open weights — RL variant)](https://huggingface.co/deepseek-ai/deepseek-math-7b-rl) — model id `deepseek-ai/deepseek-math-7b-rl`
- [Hugging Face (open weights — Instruct variant)](https://huggingface.co/deepseek-ai/deepseek-math-7b-instruct) — model id `deepseek-ai/deepseek-math-7b-instruct`
- [Hugging Face (open weights — Base variant)](https://huggingface.co/deepseek-ai/deepseek-math-7b-base) — model id `deepseek-ai/deepseek-math-7b-base`

## DeepSeek Math versions

- [DeepSeek-Math-V2](https://ai-tldr.dev/models/deepseek-math-v2/) (27 November 2025)
- DeepSeekMath (April 2024) — this page

## FAQ

### Is DeepSeekMath open source and free to use?

Yes. The weights are published on Hugging Face in three variants (Base, Instruct, and RL). The code is MIT-licensed and the model weights are covered by the DeepSeek Model License, which permits commercial use. You can download, run, and fine-tune the model yourself.

### What is GRPO, and why does DeepSeekMath matter?

GRPO (Group Relative Policy Optimization) is the reinforcement-learning algorithm introduced in the DeepSeekMath paper. It is a memory-efficient variant of PPO that removes the separate critic model and instead computes a baseline from a group of sampled answers. DeepSeek later used GRPO to train DeepSeek-R1, so DeepSeekMath is a direct technical ancestor of its reasoning models.

### How well does DeepSeekMath perform on math benchmarks?

The DeepSeekMath-RL 7B variant scores 51.7% on the competition-level MATH benchmark and 88.2% on GSM8K using chain-of-thought without external tools. With self-consistency over 64 samples, the model reaches 60.9% on MATH — approaching the MATH accuracy of Gemini-Ultra and GPT-4 despite being only 7B parameters.

### Is DeepSeekMath still the latest model in its line?

No. DeepSeekMath (2024) is the original 7B model. DeepSeek released DeepSeek-Math-V2 in November 2025 — a much larger 685B open-weight self-verifying prover. The original 7B DeepSeekMath weights are still available, but V2 is the current flagship of the DeepSeek Math line.

## Links

- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning (arXiv)](https://arxiv.org/abs/2402.03300)
- [DeepSeek-Math GitHub repository](https://github.com/deepseek-ai/DeepSeek-Math)
- [deepseek-math-7b-rl model card (Hugging Face)](https://huggingface.co/deepseek-ai/deepseek-math-7b-rl)
- [deepseek-math-7b-base model card (Hugging Face)](https://huggingface.co/deepseek-ai/deepseek-math-7b-base)
- [deepseek-math-7b-instruct model card (Hugging Face)](https://huggingface.co/deepseek-ai/deepseek-math-7b-instruct)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/deepseek-math/