DeepSeek publishes DeepSeekMath-V2 with self-verifiable reasoning

DeepSeek released DeepSeekMath-V2, a 685-billion-parameter model built on the base of DeepSeek-V3.2 and released as open weights under an Apache 2.0 licence. Rather than optimising only for correct final answers, the model was trained with a verifier-first pipeline: a separate verifier component, itself trained with reinforcement learning, scores each step of a generated proof for logical rigour and completeness, giving the model a feedback loop closer to how a mathematician drafts and revises a proof than to standard answer-matching reward.

DeepSeek reported gold-level performance on the 2025 International Mathematical Olympiad and the 2024 China Mathematical Olympiad, and a score of 118 out of 120 on the 2024 Putnam competition. The company contrasted this with general-purpose frontier models such as GPT-4 and Claude, which it said reached only silver-level IMO performance, and with Google DeepMind’s IMO-gold system, which combined self-verification with a closed release; DeepSeek positioned DeepSeekMath-V2 as the first model to combine gold-level olympiad results, built-in self-verification and full open-weight availability.

The release extended a pattern of Chinese labs, and DeepSeek in particular, using open-weight releases on narrow, rigorously scored domains — mathematics competitions chief among them — to demonstrate frontier capability without the compute or safety-evaluation overhead of a general-purpose flagship release, continuing the argument that open-weight models could match or lead on well-defined technical benchmarks even where they trailed on broader assistant tasks.