Mistral released Leanstral 1.5, an open-source formal-reasoning model with 6B active parameters designed for proof engineering and code verification.
- Saturates miniF2F benchmarks with 100% success rate
- Solves 587/672 PutnamBench problems at an estimated cost of $4 per problem
- Achieves state-of-the-art results on FATE-H (87%) and FATE-X (34%)
- Discovered 5 previously unknown bugs across 57 tested open-source repositories
- Fully open-sourced under the Apache-2.0 license with weights on Hugging Face