Skip to main content
Back

DeepSeek Releases DeepSeek-Prover-V2-7B to Write Verified Math Proofs on a Single GPU

DeepSeek releases a runnable 7B theorem prover for Lean 4 that bridges informal math reasoning with machine-verifiable formal proofs

  • DeepSeek releases DeepSeek-Prover-V2-7B, an open-source Lean 4 formal theorem prover runnable on a single GPU.
  • The 7B model scores 76-78% on MiniF2F at moderate budgets, up from 63.5% in V1.5, and adds 13 unique PutnamBench solutions the 671B missed.
  • Training uses a novel pipeline: DeepSeek-V3 decomposes theorems into subgoals, the 7B proves each subgoal, and RL with binary pass/fail feedback unifies informal and formal reasoning.
  • The flagship 671B model hits 88.9% on MiniF2F (pass@8192) and solves 49/658 PutnamBench problems, setting a new state-of-the-art.
  • A new benchmark, ProverBench, ships with 325 problems spanning AIME, calculus, algebra, analysis, and probability.
  • The 7B model requires ~16 GB VRAM in BF16; quantized GGUF versions run on 16 GB cards like the RTX 3080 Ti.

DeepSeek-Prover-V2-7B is now available on Hugging Face, giving researchers and developers a locally-runnable model that can write and verify formal mathematical proofs in Lean 4. It is the smaller sibling of the flagship 671B model, purpose-built for the same task but sized to fit on a single consumer GPU. The release comes alongside a research paper and a brand-new benchmark, making it one of the more complete drops in the formal reasoning space.

The problem nobody talks about

Most LLMs can describe a proof in plain English. Very few can write one that a computer can actually verify. Formal theorem proving requires generating syntactically valid code in a proof assistant like Lean 4, where every logical step is checked by a kernel that accepts no ambiguity. DeepSeek-Prover-V2 is not a general-purpose model. It is fine-tuned specifically to write formal proofs in Lean 4, a proof assistant used in academic mathematics, meaning it can take an intuitive explanation and turn it into a rigorous, machine-verifiable proof using Lean syntax.

The gap between "can explain a proof" and "can write a verified proof" has been the central bottleneck in this field. DeepSeek-Prover-V2 demonstrates a significant narrowing of the gap between natural-language and formal mathematical reasoning. That is the real headline here.

How it was built

A key innovation of DeepSeek-Prover-V2 lies in its unique cold-start training procedure. This process begins by prompting the powerful DeepSeek-V3 model to decompose complex mathematical theorems into a series of more manageable subgoals. Simultaneously, DeepSeek-V3 formalizes these high-level proof steps in Lean 4, effectively creating a structured sequence of sub-problems.

To handle the computationally intensive proof search for each subgoal, the researchers employed a smaller 7B parameter model. Once all the decomposed steps of a challenging problem are successfully proven, the complete step-by-step formal proof is paired with DeepSeek-V3's corresponding chain-of-thought reasoning. Think of it as a divide-and-conquer strategy: a large general model sketches the high-level plan, and a small specialist model grinds through the formal details.