Literature · AI-MO
Kimina-Prover Distill 1.7B
A 1.7-billion-parameter language model that generates Lean 4 proofs for competition-level mathematical theorems.
What Kimina-Prover Distill 1.7B is
Kimina-Prover Distill 1.7B is a specialised language model developed jointly by Project Numina and the Kimi teams under AI-MO. Designed for formal mathematical reasoning, it generates step-by-step proofs in Lean 4, an interactive theorem proving environment and programming language.
The model addresses automated theorem proving on competition-style and olympiad-level mathematical statements. By providing an automated approach to Lean 4 formalisation, it assists researchers and mathematicians working to verify mathematical statements and explore automated reasoning using accessible computational resources.
How it works
The model is built on the Qwen3-1.7B base architecture. It is produced by distilling Kimina-Prover-72B, a larger 72-billion-parameter model trained with large-scale reinforcement learning for mathematical problem solving.
During inference, the model receives a prompt that includes an informal problem description and a formal Lean 4 statement, generating a completed formal proof as text output.
Key facts
- Parameter count
- 1.7 billion
- Base architecture
- Qwen/Qwen3-1.7B
- Distilled from
- Kimina-Prover-72B
- Target formal language
- Lean 4
- GPU memory requirement
- 24 GB
- MiniF2F-test Pass@32 accuracy
- 72.95%
What you can use it for
- Lean 4 formal proof generation
- Generates formal mathematical proof steps and tactics to close goals given a theorem statement defined in Lean 4.
- Olympiad mathematics formalisation
- Assists in formalising and solving competition-level mathematical problems by translating problem statements into verified formal proofs.
- Automated reasoning benchmarking
- Serves as an evaluation baseline for automated theorem proving research on formal benchmarks such as MiniF2F.
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 Distill 1.7B
Strengths and limits
Good at
- Achieves 72.95% accuracy on the MiniF2F-test benchmark under a Pass@32 evaluation setting.
- Runs on a single consumer GPU equipped with 24 GB of memory despite tackling complex olympiad-style reasoning tasks.
- Transfers reasoning capabilities from a large reinforcement-learning-trained 72-billion-parameter parent model into a compact footprint.
Watch out for
- Targeted specifically at Lean 4 and cannot generate formal proofs for other interactive theorem provers.
- High accuracy on difficult problems relies on sampling multiple attempts, such as Pass@32, rather than a single deterministic generation.
- Requires up to 24 GB of GPU memory for serving and inference.
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 Distill 1.7B 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 Distill 1.7B 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 16GB, typically in 12s, and a run costs about 70 credits. Visitors without an account get 1,000 credits a month.
Which interactive theorem prover does this model support?
The model generates formal proofs specifically for Lean 4.
How was Kimina-Prover Distill 1.7B trained?
It was created through distillation from Kimina-Prover-72B, a model trained via large-scale reinforcement learning, using Qwen3-1.7B as the base model.
What benchmark performance does the model attain?
It attains 72.95% accuracy on the MiniF2F-test benchmark using Pass@32 sampling.
What hardware is necessary to run inference?
The model is designed to run on a single consumer GPU with at least 24 GB of GPU memory.
Written from Kimina-Prover Distill 1.7B'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.