← All models

DeepSeekMath-V2

Published Publisher DeepSeek

Overview

DeepSeekMath-V2 is a mathematical reasoning model optimized for natural-language theorem proving. It is built on DeepSeek-V3.2-Exp-Base and can generate, assess and iteratively refine proofs. Its training approach combines proof verification, meta-verification and reinforcement learning. (github.com)

Capabilities and uses

The model is intended for tasks such as mathematical problem solving, proof generation and proof review. DeepSeek reports that, with scaled test-time computation, it achieved gold-medal-level scores on IMO 2025 and CMO 2024 and scored 118/120 on Putnam 2024. These were research evaluations, not official competition medals. (arxiv.org)

Limitations

Natural-language proofs and self-verification are not guarantees of correctness; important results require independent expert review or formal verification. The strongest reported competition results used substantial search and verification computation, and the hardest IMO-level problems remained challenging. Performance should not be assumed to match those results in a single ordinary inference. (arxiv.org)

Availability and license

DeepSeek has published model weights. Its model card states that the repository and weights are licensed under Apache 2.0. The official repository refers users to DeepSeek-V3.2-Exp inference resources. (huggingface.co)

Pricing

As of November 17, 2026, DeepSeekMath-V2 is not listed on either DeepSeek's official China or international API pricing page. Its official CNY and USD input and output prices could not be verified; prices for other DeepSeek models should not be applied to it. Self-hosting may incur infrastructure costs. Pricing sources checked: China API pricing and international API pricing. (api-docs.deepseek.com)

Publication and official sources