Premium

DeepSeek’s new Math-V2 AI model can solve and self-verify complex theorems

DeepSeek-Math-V2 is said to match the performances of OpenAI and Google DeepMind’s models on problems from the International Maths Olympiad 2025.

DeepSeek launched R1-0528 on developer platform Hugging Face, but has yet to make an official public announcement. It did not publish a description of the model or comparisons. (Express Photo)

Chinese AI startup DeepSeek on Thursday, November 27, unveiled a new open-weight AI model designed to generate and self-verify mathematical theorems using advanced reasoning skills that the company says were developed specifically for this task.

Named DeepSeek-Math-V2, the specialist mathematical reasoning LLM (large language model) is said to possess strong theorem-proving capabilities. The model weights are publicly available for download under the Apache 2.0 open-source license on platforms such as Hugging Face and GitHub.

The mathematical reasoning model has been built on top of DeepSeek-V3.2-Exp, an experimental AI model introduced by the Huangzhou-based startup in September this year.

DeepSeek-Math-V2 focuses on self-verifiable mathematical reasoning. It comprises two key components: a verifier that checks mathematical proofs step-by-step, and a theorem generator that is capable of fixing its own mistakes.

“By scaling reasoning with reinforcement learning that rewards correct final answers, LLMs have improved from poor performance to saturating quantitative reasoning competitions like AIME and HMMT in one year. However, this approach faces fundamental limitations,” DeepSeek said in a technical paper accompanying the launch of the model.

Also Read | Microsoft unveils Fara-7B, a lightweight AI model that can use a PC from a single screenshot

According to DeepSeek, pursuing higher final answer accuracy does not guarantee correct reasoning and is inapplicable for mathematical tasks like theorem proving which requires rigorous step-by-step derivation. Instead, DeepSeek-Math-V2 has been designed to solve open problems without known solutions by using self-verification as a way of scaling test-time compute and performing deeper reasoning.

In terms of performance, DeepSeek said that Math-V2 achieved gold medal-worthy scores when tested on math problems from the International Mathematical Olympiad (IMO) 2025 and CREST Mathematics Olympiad (CMO) 2024. It also reportedly achieved a high score of 118 out of 120 on math problems from the Putnam 2024 mathematical competition.

Story continues below this ad

“While much work remains, these results suggest that self-verifiable mathematical reasoning is a feasible research direction that may help develop more capable mathematical AI systems,” DeepSeek said.

DeepSeek’s gold medal-level score on IMO 2025 problems puts it in the elite company of OpenAI and Google DeepMind, whose unreleased models posted similar results at the prestigious math competition held earlier this year.

Also Read | What is Antigravity, Google’s new agentic AI coding platform raising fresh security concerns?

However, the key distinction is that this was the first year IMO organisers formally admitted AI models in the competition, and while Google was part of that inaugural cohort, OpenAI and DeepSeek were not included.

While LLMs are adept at churning out vast amounts of text, the recent progress made by reasoning models in the mathematical domain could aid researchers in cracking long-standing research problems related to fields like cryptography and space exploration.

Story continues below this ad

DeepSeek’s Math-V2 model also marks a significant step forward for the open ecosystem, which has come to be dominated by China. A recent study by the Massachusetts Institute of Technology (MIT) and open-source AI start-up Hugging Face found that the total share of downloads of new Chinese-made open models rose to 17 per cent in the past year, which may have given the country a crucial edge over the United States in the global market for open AI models.