DeepSeek’s Math‑V2 Delivers IMO‑Level Mathematical Reasoning

DeepSeek has released DeepSeek‑Math‑V2 (Math‑V2), an open‑source Mixture‑of‑Experts reasoner that achieves gold‑medal performance on the International Mathematical Olympiad (IMO). According to The Rundown, the model solved five of six IMO 2025 problems and scored 118/120 on the 2024 Putnam — a result that surpasses the top human Putnam score. On the IMO ProofBench benchmark Math‑V2 hit 61.9%, nearly matching Google’s specialized Gemini Deep Think and far outperforming GPT‑5’s ~20% result.

Math‑V2 uses a generator–verifier architecture: one model proposes proofs and another critiques them. The verifier assigns confidence scores to individual steps, forcing the generator to refine weak logic and enabling step‑by‑step self‑debugging rather than rewarding only final answers. This approach yields stronger, more reliable mathematical reasoning.

By open‑sourcing a system that rivals internal heavyweight models, DeepSeek has effectively democratized research‑level symbolic reasoning. The newsletter frames this as breaking a prior monopoly on frontier mathematical reasoning and suggests the technique could be a blueprint for building agents that debug their own thought processes — with potential impact in high‑stakes domains like engineering where correctness matters.