Mistral Leanstral 1.5 Formal Verification Proof Pipeline

Mistral Leanstral 1.5 (released July 2, 2026, Apache-2.0) is a 119B parameter mixture-of-experts model with 6.5B active parameters specialized for Lean 4 formal verification. It saturates the miniF2F benchmark at 100%, solves 587 of 672 PutnamBench problems at approximately $4 per problem versus $300+ for Seed-Prover, and achieves state-of-the-art 87% on FATE-H and 34% on FATE-X benchmarks. The model features 128 experts with 4 active per token, a 256K token context window, and multimodal input capabilities. Leanstral 1.5 operates through Mistral Vibe, a proof engineering interface for Lean 4 that provides tactic generation, proof completion, and autoformalization. Beyond mathematical theorem proving, the model has demonstrated the ability to catch real bugs in Rust code by formally verifying safety properties. The Apache-2.0 license makes it freely deployable for commercial use, fine-tuning, and on-premise inference.

BUSINESS PROBLEM

Formal verification has remained inaccessible to most development teams because it requires specialized PhD-level expertise in mathematical logic, type theory, and proof assistants. According to Mistral's benchmarks, traditional formal verification of a single non-trivial theorem costs $300+ in compute for Seed-Prover and requires hours of expert time per proof. Software bugs cost the global economy an estimated $2.4 trillion annually according to a 2025 Cambridge University study on software failure costs. Industries such as fintech, medical devices, aerospace, and smart contracts face disproportionate risk because bugs in these systems cause financial loss, safety hazards, or legal liability. Leanstral 1.5 bridges the gap between the need for formally verified software and the scarcity of formal methods expertise by making proof engineering accessible through natural language interaction at commodity compute costs.

WHO BENEFITS

Security engineer at a DeFi protocol company who needs formally verified smart contracts to prevent multi-million-dollar exploits without hiring a PhD-level formal methods specialist. Compiler engineer at a chip design firm who needs to verify correctness properties of Rust-based firmware for safety-critical embedded systems. Research scientist at a university AI lab who uses Lean 4 for mathematical formalization and wants to automate proof generation for large-scale theorem libraries.

HOW IT WORKS

Step 1 - Access Setup. Use the free Mistral Vibe interface at vibe.mistral.ai or download weights from HuggingFace (Apache-2.0). Step 2 - Problem Input. Provide a theorem statement in Lean 4 syntax or describe it in natural language for autoformalization. Step 3 - Tactic Generation. Leanstral 1.5 generates candidate tactics and proof steps using its 128-expert MoE architecture. Step 4 - Proof Exploration. The model explores multiple proof paths, backtracking when a tactic fails. Step 5 - Verification. Each completed proof is checked by the Lean 4 kernel, providing absolute certainty of correctness. Step 6 - Integration. Verified proofs are committed to the codebase alongside the implementation they verify. Step 7 - CI/CD Pipeline. For ongoing verification, Leanstral 1.5 runs as part of a CI pipeline, catching regressions when code changes break formal properties.

TOOL INTEGRATION

Leanstral 1.5 - Apache-2.0 Lean 4 theorem-proving model (119B MoE, 6.5B active). Mistral Vibe - Proof engineering interface for Lean 4. Lean 4 - Programming language and proof assistant. PutnamBench - Benchmark of 672 formalized Putnam Competition problems. miniF2F - Formal-to-formal mathematics benchmark. FATE-H/FATE-X - Formalization benchmarks. HuggingFace - Model weights distribution under Apache-2.0 license.

ROI METRICS

Proof cost reduced from $300+ per problem to ~$4 per problem (vs Seed-Prover). 100% on miniF2F benchmark - saturates the benchmark entirely. 587/672 PutnamBench problems solved - 87.4% success rate. Real Rust bugs detected in production code through formal property verification. Apache-2.0 license eliminates per-seat licensing costs for commercial deployments.

CAVEATS

MEDIUM - Lean 4 expertise is still required to formulate the properties to prove; Leanstral handles the proof search but not the specification. LOW - The model specializes in Lean 4; proofs in other systems (Coq, Isabelle, Agda) are not supported. MEDIUM - Proof generation for very complex theorems may require multiple attempts; success is not guaranteed for all conjectures. LOW - Mistral Vibe is the recommended interface; CLI and API usage require additional setup for CI/CD integration.