News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
DeepSeekMath

DeepSeekMath

Open language model for mathematical reasoning

DeepSeekMath is a 7-billion parameter language model built by DeepSeek AI for advanced mathematical problem solving. Released in early 2024, it continues pre-training on a large corpus of math-related web data. The model achieves high scores on competition-level benchmarks without external tools, matching performance of larger proprietary systems.

Share on WhatsApp

  • Scores 51.7% on the MATH benchmark without tools
  • Trained on 120 billion math-related tokens
  • Uses Group Relative Policy Optimization for training
  • Supports commercial use under its model license
  • Based on DeepSeek-Coder-Base-v1.5 7B architecture

What it is

DeepSeekMath is a specialised language model with 7 billion parameters designed to handle complex mathematical reasoning tasks. Developed by DeepSeek AI, it was introduced in February 2024 as an open-weight model. It builds upon the DeepSeek-Coder-Base-v1.5 architecture but focuses specifically on mathematics. The model processes text inputs and generates step-by-step solutions, aiming to solve competition-level problems without needing external calculators or code interpreters.

How it came about

The model was announced in a paper submitted to arXiv in February 2024, with revisions following in April. It extends the work done on DeepSeek-Coder-Base-v1.5 by continuing its pre-training phase. The development team, led by Zhihong Shao and colleagues, focused on harnessing publicly available web data. They engineered a specific pipeline to select high-quality mathematical content from Common Crawl to improve the model's reasoning capabilities.

Inside the model

The architecture is a dense decoder-only transformer. Training involved 120 billion tokens related to mathematics, sourced from Common Crawl alongside natural language and code data. A key innovation is the use of Group Relative Policy Optimization (GRPO), a variant of Proximal Policy Optimization. This method enhances reasoning skills while reducing memory usage during training. The tokenizer automatically adds a beginning-of-sequence token, and the model does not support system prompts in its current version.

What it can do

DeepSeekMath scores 51.7 per cent on the MATH benchmark, which tests competition-level mathematics. This result was achieved without using external toolkits or voting techniques. When using self-consistency over 64 samples, the score rises to 60.9 per cent. The developers state this performance approaches that of Gemini-Ultra and GPT-4. The model supports both English and Chinese queries, requiring chain-of-thought prompting for best results.

How to use it

The model weights are available openly under a Model License that permits commercial use. The code repository uses the MIT License. Users can download the model from platforms like Hugging Face. Interaction requires specific prompting formats, such as asking the model to reason step by step and place the final answer in a boxed format. No API-only restriction is mentioned, allowing local deployment.

Questions people ask about DeepSeekMath

What is DeepSeekMath used for?

It is designed for solving complex mathematical problems, including competition-level questions. It generates step-by-step reasoning in text form without relying on external code executors or calculators.

Is DeepSeekMath free to use commercially?

Yes, the model weights are released under a Model License that explicitly supports commercial use. The accompanying code repository is licensed under the permissive MIT License.

How large is the DeepSeekMath model?

It has 7 billion parameters. It is based on the DeepSeek-Coder-Base-v1.5 7B model, which was further pre-trained on a large dataset of mathematical content.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on DeepSeekMath yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2402.03300, huggingface.co/deepseek-ai/deepseek-math-7b-instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.