DeepSeekMath-V2: open-weights self-verifying prover reaches IMO 2025 gold level and 118/120 on Putnam 2024
Confirmed
Importance: major (4 of 5)The takeaway
DeepSeek released DeepSeekMath-V2 (685B parameters, built on DeepSeek-V3.2-Exp-Base, Apache 2.0).
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 29 September 2026
Your AI and this story
- GPT-6 AstraIn its training data
- Claude Opus 5.5In its training data
- Gemini 4 HydrogenIn its training data
- Grok 4.7In its training data
It happened before the cutoffs of all four assistants, so it can be in their training data.
Key facts
- Paper: arXiv 2511.22570 ‘DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning’ (27 Nov 2025)
- 685B parameters; base DeepSeek-V3.2-Exp-Base; Apache 2.0 weights on Hugging Face
- Gold-level scores on IMO 2025 and CMO 2024; 118/120 on Putnam 2024 (scaled test-time compute)
- Method: faithful LLM proof verifier plus meta-verification to cut hallucinated issues; the generator is rewarded for finding and fixing its own errors; verifier compute is scaled to auto-label hard proofs without human annotation
What happened
Four months after closed models from Google DeepMind and OpenAI reached IMO gold, DeepSeek released open weights for a proof-writing model at the same level. It made self-verification (generator + verifier + meta-verifier) the main training signal instead of final-answer rewards.
Why it matters
It made olympiad-level natural-language proof generation reproducible outside the big US labs. The generate-then-verify recipe became a common pattern in 2026 AI-for-math systems.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 2 primary
Primary
- arXiv 2511.22570arxiv.org, paper
- Hugging Face: deepseek-ai/DeepSeek-Math-V2huggingface.co, code
Changes
- Filed