DeepSeek Releases DeepSeek-Math-V2 with Lean 4 Formal Proof Verification and 91.4% IMO Accuracy

DeepSeek has officially released DeepSeek-Math-V2, an advanced open-weights reasoning model purpose-built for rigorous mathematical problem solving, formal theorem proving, and interactive verification within the Lean 4 proof assistant. The release establishes a new milestone for open-source mathematical AI, achieving 91.4% accuracy on IMO-level competition benchmarks — a result that matches or exceeds several proprietary frontier models on the most demanding mathematical reasoning tasks.

The Lean 4 Integration

DeepSeek-Math-V2's most significant architectural innovation is its native integration with Lean 4, a functional programming language and formal mathematical proof assistant used by professional mathematicians and computer scientists. The model can generate Lean 4 proof code from natural language problem statements, submit the proofs to the Lean 4 kernel for formal verification, and iteratively refine failed proofs based on the kernel's error output. This creates a closed feedback loop between neural reasoning and formal symbolic verification that prevents hallucinated or logically incorrect proofs from being accepted as valid.

Benchmark Performance

• IMO Problem Set (Competition Level): 91.4% accuracy
• MATH-500 benchmark: 96.7% accuracy
• AIME 2024: 83.3% solve rate (10/12 problems)
• MathBench (Chinese High School Advanced): 98.2%
• Lean 4 formal verification: 78.3% of proofs pass automated verification on first attempt

Chain-of-Thought Formal Verification Pipeline

DeepSeek-Math-V2 introduces what the team calls a Chain-of-Thought Formal Verification pipeline. For each mathematical problem, the model first generates a natural language solution with step-by-step reasoning. It then automatically translates each reasoning step into a corresponding Lean 4 proposition and checks each step formally. Steps that fail formal verification are flagged and the model re-reasons from that point — effectively catching logical errors that would be invisible in natural language text alone. This pipeline is activated automatically when the model is invoked with `verification_mode=True` in the API.

Open-Weights Access and Research Impact

DeepSeek-Math-V2 is released as fully open weights under a research license permitting commercial use for organizations under 100 employees. The weights are available on Hugging Face in 7B and 36B parameter variants. The 36B model achieves the headline benchmark results while the 7B model is optimized for local inference on consumer-grade hardware (M2 MacBook Pro, RTX 4090). For academic researchers, mathematics PhD students, and competition mathematics communities, this represents the most capable freely available mathematical reasoning model to date.

Applications Beyond Competition Math

Beyond competition mathematics, the formal verification pipeline has immediate applications in software verification (proving program correctness), cryptography (verifying cryptographic protocol soundness), hardware design verification (checking chip logic against formal specifications), and financial derivative pricing models (formally verifying risk calculation formulae). DeepSeek Math-V2 is available via DeepSeek's API and HuggingFace at `deepseek-ai/DeepSeek-Math-V2-Instruct`.