DeepSeek made waves in early 2025, launching one of the world"s first free-to-access thinking models. Now, the Chinese firm has just released DeepSeekMath-V2 with the objective of achieving self-verifiable mathematical reasoning and rigorous step-by-step derivation, with a focus on theorem proving, moving beyond reliance on correct final numerical answers.
The company said that its new model uses a generation-verification loop. It said that an accurate LLM-based verifier has been trained for theorem proving. DeepSeek also trained a proof generator using the verifier as the reward model. The proof generator is incentivized to identify and resolve issues in its own proofs, and verification scaling is used to automatically label new, hard-to-verify proofs, providing training data to continuously improve the verifier.
DeepSeekMath-V2 has already demonstrated strong theorem-proving abilities in recent math competitions. It achieved gold-level scores on IMO 2025 and CMO 2024, as well as a near-perfect 118/120 on Putnam 2024 with scaled test-time compute.
The new model is built on DeepSeek-V3.2-Exp-Base, and is available on HuggingFace. For inference support, DeepSeek recommends checking out the support in the DeepSeek-V3.2-Exp GitHub repository.
The launch of this model is pretty interesting and could help unlock a better understanding of mathematics, which could lead to new scientific discoveries to improve healthcare and technology. AI firms won"t stop here; they will continue to find new methods to increase mathematical understanding, and open access to this model will only help to accelerate improvements. Don"t expect proofs of the Millennium Problems to be uncovered just yet, but it is important to work towards helping people figure those out.
You can find out more about the model on its dedicated GitHub page. There is a research paper about it that can be downloaded as a PDF. You can also find the model for download over on HuggingFace.