Mistral has just released Leanstral 1.5, a Large Language Model designed for formal verification and licensed under Apache 2.0. The most striking figure isn’t just its benchmark-topping performance, but the architecture: 119 billion total parameters, of which only 6 billion are active per inference. This ratio effectively redraws the boundaries of on-premise deployment for proof engineering tools, because it enables local execution of workloads that until recently would have required dedicated clusters.

The model essentially saturates the miniF2F benchmark (a test of mathematical theorems), solves 587 out of 672 PutnamBench problems – the notoriously difficult North American undergraduate competition – and sets new state-of-the-art references on FATE-H (87%) and FATE-X (34%), two metrics designed to evaluate real-world code verification capabilities. And this isn’t academic exercise: in tests across 57 open source repositories, Leanstral 1.5 uncovered 5 previously unknown bugs, proving its concrete effectiveness in validating software specifications and implementation correctness.

Training followed a three-stage path: mid-training on domain-specific corpora, supervised fine-tuning, and reinforcement learning via the CISPO algorithm (Contextual Inference with Self-Play Optimization, a variant tailored to formal domains). The result is a model that excels not only in automated theorem proving but also in “agentic proof engineering,” where the system actively explores proof spaces much like a human mathematician would.

What does this mean for those considering on-premise deployment? The key advantage lies in the low active parameter count. Six billion is a threshold that today can be handled by a single consumer GPU or a well-specced workstation, without needing cloud services. For teams working on proprietary code, encrypted software, or in environments with data residency constraints, the ability to run a formal verifier without ever exposing source code externally is a game-changer. The Apache 2.0 license further removes legal uncertainty and allows unrestricted forking, modifications, and integrations.

Mistral has not yet released precise hardware requirements, but the resource profile puts Leanstral 1.5 close to models already being served on single GPUs with 16–24 GB of VRAM, potentially using quantization if an even smaller footprint is needed. Perhaps the most compelling aspect is that the entire verification workflow remains confined to the corporate infrastructure, aligning with security policies that make the cloud only partially viable. In an era where data sovereignty is no longer a talking point but a contractual requirement, models like this mark a tangible shift for development tooling.