Curiosity · AI Model

DeepSeek-Prover V2

DeepSeek-Prover V2 is DeepSeek's second-generation formal theorem prover for Lean 4, released April 2025. Building on DeepSeek-Math and DeepSeek-Prover V1, it couples an RL-trained proof generator with a Lean verifier, a self-play loop, and supervised warm-start on millions of auto-formalised proofs. V2 reaches new state-of-the-art on the MiniF2F competition-math benchmark and is notable for being a fully open alternative to DeepMind's AlphaProof, with weights released on Hugging Face.

Model specs

Vendor
DeepSeek
Family
DeepSeek-Prover
Released
2025-04
Context window
32,768 tokens
Modalities
text, code

Strengths

  • Fully open — weights, code, and training recipes
  • SOTA on MiniF2F among open systems
  • Strong self-play + RL training pipeline
  • Integrates cleanly with Lean 4 tooling

Limitations

  • Proof search is compute-heavy
  • Behind AlphaProof on IMO-level problems
  • Lean-only output — no natural-language explanations
  • Requires Lean 4 expertise to use effectively

Use cases

  • Open research on theorem proving
  • Auto-formalisation of informal math proofs
  • Competition math (Putnam / IMO-style) solvers
  • Baseline for RL-over-Lean methods

Benchmarks

BenchmarkScoreAs of
MiniF2F (test)new SOTA at release2025-04
PutnamBenchbest open system2025-04

Frequently asked questions

What is DeepSeek-Prover V2?

An open-weights formal theorem prover for Lean 4 from DeepSeek, trained with RL and self-play on auto-formalised proof data.

How does it compare to AlphaProof?

AlphaProof is stronger on IMO-level problems but closed. DeepSeek-Prover V2 is the leading open alternative and holds SOTA on MiniF2F among public systems.

What licence is DeepSeek-Prover V2 under?

Weights are released under the DeepSeek community licence; code is open-source on GitHub.

Sources

  1. DeepSeek-Prover V2 on Hugging Face — accessed 2026-04-20
  2. DeepSeek-Prover V2 paper (arXiv) — accessed 2026-04-20