Literature · AI-MO
Kimina-Prover Distill 8B
A distilled 8-billion-parameter language model for automated Lean 4 theorem proving in mathematics.
What Kimina-Prover Distill 8B is
Kimina-Prover Distill 8B is an open-weights formal theorem-proving model developed jointly by Project Numina and the Kimi team (AI-MO). Designed to write formal proofs in the Lean 4 interactive theorem prover, it specialises in competition-style and olympiad-level mathematical problems.
The model is a distilled variant of Kimina-Prover-72B, transferring the mathematical reasoning capabilities of the larger system into an 8-billion-parameter footprint. It allows researchers to generate candidate Lean 4 proof traces from paired natural language problem descriptions and formal statements.
On the MiniF2F-test benchmark, it achieves 77.86% accuracy under Pass@32 evaluation, demonstrating practical utility for automated formalisation and computer-verified mathematical proof search.
How it works
Kimina-Prover Distill 8B is fine-tuned from Qwen/Qwen3-8B. It is produced by distilling Kimina-Prover-72B, a larger model trained through large-scale reinforcement learning over proof traces.
The model operates as a text-generation chat model. It takes prompts containing an informal problem description alongside a formal Lean 4 theorem declaration, generating step-by-step tactics and proof scripts to complete the proof.
Key facts
- Base architecture
- Qwen3-8B
- Parameter count
- 8 billion
- Formal proof language
- Lean 4
- Teacher model
- Kimina-Prover-72B
- Benchmark accuracy
- 77.86% Pass@32 on MiniF2F-test
- Recommended GPU memory
- 40 GB
What you can use it for
- Automated Lean 4 theorem proving
- Generates formal tactics and proofs for competition-style mathematical theorems formulated in Lean 4.
- Olympiad-level mathematical reasoning
- Solves challenging mathematical contest problems by reasoning step by step across algebra, geometry, and number theory in a verifiable syntax.
- Assisted formalisation
- Assists researchers in completing proof goals within Lean 4 environments, using formal statements combined with natural language context.
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 8B
Strengths and limits
Good at
- Attains 77.86% Pass@32 accuracy on the MiniF2F-test formal mathematics benchmark.
- Distils capabilities from a 72-billion-parameter reinforcement-learned model into an 8-billion-parameter model.
- Native support for Lean 4 mathematical libraries including Mathlib and Aesop.
Watch out for
- Demands significant compute resources, requiring 40 GB of GPU memory for execution.
- Performance is evaluated using multi-sample Pass@32 search rather than single-pass deterministic generation.
- Targeted specifically at Lean 4 and competition mathematics, without documented coverage of other formal languages.
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 8B 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 8B 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.
Which formal verification language does this model target?
It is designed to generate formal proof traces in Lean 4, relying on libraries such as Mathlib and Aesop.
What model was used as the base for Kimina-Prover Distill 8B?
It is fine-tuned from Qwen/Qwen3-8B and distilled from the larger Kimina-Prover-72B model.
What input format does the model expect?
It accepts conversational prompts containing a natural language description of the problem alongside a formal Lean 4 theorem declaration.
How well does the model perform on standard benchmarks?
The model achieves an accuracy of 77.86% under Pass@32 sampling on the MiniF2F-test benchmark.
Written from Kimina-Prover Distill 8B'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.