Leanstral: Open-source agent for trustworthy coding and formal proof engineering

On March 16, 2026, Mistral AI released Leanstral, the first open-source AI model built specifically for Lean 4 — a programming language that forces you to mathematically prove your code works before it runs (Source 2). Think of Lean 4 like a very strict editor that won't let you save a document until every claim you make has a cited source. Except instead of citations, you're providing mathematical proofs that your code does exactly what you say it does (Source 3).

Leanstral is tiny compared to other AI coding tools: only 6.5 billion parameters active at once (out of 119 billion total), yet it beat much larger models on realistic proof-engineering tasks (Source 3). For context, that's like having a compact car that outperforms SUVs because it was designed specifically for city driving, not off-roading.

Mistral released everything under Apache 2.0 license (fully open for anyone to use commercially), with free API access and integration into their "Mistral Vibe" platform using the /leanstral command (Source 3). They also released FLTEval, a new test that checks how well AI can complete real mathematical proofs — like work on Fermat's Last Theorem — instead of just solving textbook problems (Source 2).

How Leanstral performed: On a single attempt, Leanstral scored 26.3 on FLTEval. The closest open-source competitor, Qwen3.5 (with 17 billion active parameters), needed 4 attempts to reach 25.4 (Source 2). Leanstral also beat Claude Sonnet 4.6 (which scored 23.7) at a fraction of the cost — approximately $36 for two passes versus significantly more for Claude (Source 3).

The model handles 256,000 tokens of context (Mistral recommends 200k for best results) and works with text, images, and 11 languages including English, Chinese, Japanese, and Arabic (Source 3). It was specifically trained to work with lean-lsp-mcp, a commonly used tool in the Lean ecosystem (Source 2).

The real story here is trust. When an AI writes code today, you run tests and hope you caught the edge cases. When Leanstral writes code in Lean 4, the code won't even compile unless there's a mathematical proof attached showing it behaves correctly (Source 6). That's the difference between "the unit tests passed" and "this is mathematically guaranteed to never divide by zero" (Source 6).

This matters most in high-stakes domains where bugs kill people or cost billions: aerospace software, financial trading systems, cryptography, medical devices, self-driving cars (Source 3). Right now, those industries spend enormous time manually reviewing AI-generated code because they can't afford mistakes. Leanstral shifts the bottleneck: instead of humans reviewing AI's work, humans write specifications and the AI proves its implementation matches. As one commenter noted, "verification suites accrue over time into a detailed repository that puts zero tokens in context when code is correct" (Source 4). In other words: once proven correct, it doesn't need constant re-checking.

Here's the uncomfortable truth: Most developers will never touch Lean 4 directly — it has a steep learning curve and only shines where correctness is non-negotiable. But the category Leanstral creates — "verified software engineering at AI speed" (Source 5) — points toward a future where your banking app or the software updating your car's brakes can be mathematically proven safe, not just "probably fine." The question is whether formal verification tooling becomes accessible enough for everyday use, or stays locked in research labs and defense contractors.