AI-MO/Kimina-Prover-72B

Hugging Face
TEXT GENERATIONConcurrency Cost:4Model Size:72.7BQuant:FP8Ctx Length:32kPublished:Jul 9, 2025License:mitArchitecture:Transformer0.0K Open Weights Warm

AI-MO/Kimina-Prover-72B is a 72.7 billion parameter theorem proving model developed by Project Numina and Kimi teams, based on the Qwen2.5-72B architecture. It is specifically trained via large-scale reinforcement learning to excel at competition-style problem solving in Lean 4. This model achieves 84.0% accuracy with Pass@32 on the MiniF2F-test, making it highly effective for formal mathematics and automated theorem proving tasks.

Loading preview...

Kimina-Prover-72B: A Specialized Theorem Proving Model

Kimina-Prover-72B, developed by Project Numina and Kimi teams, is a 72.7 billion parameter model built upon the Qwen2.5-72B base architecture. Its primary focus is on advanced theorem proving, particularly for competition-style problems within the Lean 4 formal proof system. The model's development involved extensive large-scale reinforcement learning, optimizing its ability to generate and verify mathematical proofs.

Key Capabilities

  • Formal Theorem Proving: Specialized in generating proofs and solving problems in Lean 4.
  • High Accuracy: Achieves a notable 84.0% accuracy with Pass@32 on the MiniF2F-test benchmark, indicating strong performance in complex mathematical reasoning.
  • Reinforcement Learning Optimized: Benefits from large-scale reinforcement learning training, enhancing its problem-solving strategies.

Good For

  • Automated Theorem Proving: Ideal for tasks requiring automated generation or verification of mathematical proofs.
  • Lean 4 Development: Particularly useful for researchers and developers working with the Lean 4 proof assistant.
  • Mathematical Problem Solving: Excels in competition-style mathematical challenges that demand rigorous logical deduction.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p