Skip to content
← Models
Open weightsDeepSeek/10 trustVerified

DeepSeek-Prover-V2

671B open theorem prover — subgoal decomposition + RL for Lean 4 proofs, MIT licensed.

Visit official page →

About

DeepSeek-Prover-V2 (released April 30, 2025) is DeepSeek's formal-mathematics model family for generating and verifying proofs in Lean 4. The headline 671B MoE (built on DeepSeek-V3-Base) and a practical 7B variant (built on Prover-V1.5-Base, 32K context) share a key innovation: subgoal decomposition — DeepSeek-V3 breaks theorems into subgoals during training, the prover solves each, and the traces become cold-start data for RL with binary feedback from Lean's kernel verifier.

The 671B model hit 88.9% on MiniF2F and set records on PutnamBench at release; the 7B runs on a 16GB consumer GPU (or ~5GB quantized) making formal proving accessible. Released under the MIT license on Hugging Face.

Honest limits: the 671B needs datacenter hardware (~650GB weights); headline benchmarks are DeepSeek's own; it proves in Lean 4 only — informal math reasoning is not its strength; later open provers (Goedel-Prover-V2, Kimina-Prover) match or beat it far more efficiently.

Specs

Context window

Pricing

Free quota

API access

Weights

Open weights

Released

April 30, 2025

Benchmarks

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

maththeorem-provingreasoningresearcheducation

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 55; Trending score 75/100 from researcher assessment.

More in the deepseek family

Sources