Literature · AI-MO
Kimina-Prover Preview Distill 7B
A distilled 7B language model that generates formal mathematical proofs in Lean 4.
What Kimina-Prover Preview Distill 7B is
Kimina-Prover Preview Distill 7B is a 7-billion-parameter text-generation model designed for automated formal theorem proving. Developed jointly by the Project Numina and Kimi teams under AI-MO, the model generates formal proofs in Lean 4, an interactive theorem prover and programming language widely used in formalised mathematics.
The model addresses the challenge of automated reasoning on competition-style and olympiad-level mathematical problems. It is produced by distilling Kimina-Prover-Preview, an upstream model trained with large-scale reinforcement learning, onto a smaller architecture to enable formal proof generation under modest compute budgets.
How it works
The model is a distilled version of Kimina-Prover-Preview, which was trained using large-scale reinforcement learning for formal mathematical reasoning. It is based on the Qwen2.5-Math-7B-Instruct architecture and operates within the Qwen2 transformer family.
For inference, the model processes conversational prompts containing informal natural-language problem descriptions alongside formal Lean 4 statements. It then generates reasoning steps followed by complete Lean 4 proof completions.
Key facts
- Base model
- Qwen/Qwen2.5-Math-7B-Instruct
- Architecture
- Qwen2
- Parameter count
- 7B
- Primary language
- Lean 4 and English
- Output modality
- Text
- GPU memory requirement
- 40 GB
What you can use it for
- Olympiad-style Lean 4 proof generation
- Researchers can provide an informal mathematical problem and a formal Lean 4 statement to produce step-by-step formal proofs for competition-level problems.
- Undergraduate competition mathematics
- The model can be applied to solve formalised contest problems, having achieved the top position on the PutnamBench leaderboard within a small compute budget.
- Formal theorem proving benchmarking
- It serves as a benchmark model for formal mathematical reasoning evaluations, including the MiniF2F-test benchmark.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
Kimina-Prover Preview Distill 7B
Strengths and limits
Good at
- Delivers leading results on the MiniF2F-test benchmark for its parameter size and compute budget.
- Ranked first on the PutnamBench leaderboard with a small compute budget at the time of release.
- Distilled into a compact 7B parameter footprint, making it less resource-intensive than its reinforcement-learning-trained parent model.
Watch out for
- Requires high GPU resources (40 GB memory) for execution despite being a distilled 7B model.
- Input and output behaviour must be checked carefully against formal specifications, as default repository pipelines are derived from generic tags rather than bespoke harnesses.
Inputs and outputs
You provide
Prompt
The model continues from here.
You get back
The model's completion.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use Kimina-Prover Preview Distill 7B commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run Kimina-Prover Preview Distill 7B without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 12s, and a run costs about 210 credits. Visitors without an account get 1,000 credits a month.
What formal language does this model support?
The model generates formal mathematical proofs written in Lean 4, typically relying on libraries like Mathlib and Aesop.
What was the model distilled from?
It is distilled from Kimina-Prover-Preview, a larger model developed with large-scale reinforcement learning for formal mathematical reasoning.
What base architecture does the model use?
It is built by fine-tuning and distilling onto the Qwen2.5-Math-7B-Instruct base model.
How to cite it
Wang, Haiming and Unsal, Mert and Lin, Xiaohan and Baksys, Mantas and Liu, Junqi and Santos, Marco Dos and Sung, Flood and Vinyes, Marina and Ying, Zhenzhe and Zhu, Zekai and Lu, Jianqiao and Saxcé, Hugues de and Bailey, Bolton and Song, Chendong and Xiao, Chenjun and Zhang, Dehao and Zhang, Ebony and Pu, Frederick and Zhu, Han and Liu, Jiawei and Bayer, Jonas and Michel, Julien and Yu, Longhui and Dreyfus-Schmidt, Léo and Tunstall, Lewis and Pagani, Luigi and Machado, Moreira and Bourigault, Pauline and Wang, Ran and Polu, Stanislas and Barroyer, Thibaut and Li, Wen-Ding and Niu, Yazhe and Fleureau, Yann and Hu, Yangyang and Yu, Zhouliang and Wang, Zihan and Yang, Zhilin and Liu, Zhengying and Li, Jia. Kimina-Prover Preview: Towards Large Formal Reasoning Models with Reinforcement Learning. 2025. arXiv:2504.11354.
Written from Kimina-Prover Preview Distill 7B's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.