DeepSeekMath-V2

What it is

DeepSeekMath-V2 is a mathematical reasoning model from DeepSeek, released on 27 November 2025 ^[2]. Wikipedia lists it under the Apache 2.0 license ^[2]. The model is built on top of DeepSeek-V3.2-Exp-Base ^[1]. Its aim is self-verifiable mathematical reasoning, meaning that the model proves theorems rigorously rather than only producing final answers that happen to be correct ^[1].

Origins

The model is the successor to DeepSeekMath, released in February 2024, and to DeepSeek-Prover-V2, released in April 2025 ^[2].

How it works

The method rests on a verifier, a model that checks whether a proof is correct. DeepSeek trained an LLM-based verifier for theorem proving and used it as the reward model for a proof generator ^[1]. A reward model gives a score to an output, and the generator adjusts its behavior to earn higher scores. In this setup the generator is rewarded for finding and fixing issues in its own proofs before it finalizes them ^[1]. In practice this means the generator writes a proof, checks each step, repairs what it finds, and only then gives its answer.

Verification is hard to automate for new problems, and the verifier has to keep up as the generator improves. To handle this, DeepSeek scaled up verification compute so that the system could automatically label new, hard-to-verify proofs as training data for the verifier ^[1]. Verification compute here means the extra computation spent on checking proofs.

Results

DeepSeek reported gold-level scores on IMO 2025 and CMO 2024 ^[1]. The International Mathematical Olympiad and the China Mathematical Olympiad are competitions for high-school students, and gold-level results mean scores at the top tier of those competitions. On Putnam 2024, a university-level competition, the model scored 118 out of 120 with scaled test-time compute ^[1]. Test-time compute is the extra computation spent when the model is answering a question.

Licensing and openness

The DeepSeekMath-V2 repository is released under the Apache-2.0 license, but use of the model itself is subject to a separate model license ^[1]. DeepSeek also published the model's predictions for the competition problems so that outside readers could check them ^[1].

Position in the work

DeepSeekMath-V2 marks a shift in the DeepSeek math line from producing answers toward checking proofs. Its verifier-and-generator design shows how a model can be trained to judge mathematical arguments, not only to produce them ^[1].