Deepseek Prover V2 671B

DeepSeek Prover V2 671B is a 671-billion parameter Mixture-of-Experts model specialized for formal mathematical reasoning and theorem proving in Lean 4. It uses a cold-start procedure in which DeepSeek V3 decomposes complex problems into subgoals, then combines reinforcement learning with subgoal decomposition to narrow the gap between informal and formal reasoning. The model reaches an 88.9% pass ratio on the MiniF2F-test benchmark and solves 49 of 658 problems on PutnamBench, demonstrating strong capability in translating mathematical reasoning into formal proofs. It shares the architecture of DeepSeek V3 Base with supervised fine-tuning optimized for proof generation. Released on April 30, 2025, DeepSeek Prover V2 671B is open-weight under the MIT license, making it valuable for researchers and developers working with formal mathematics and theorem proving.