Mistral AI releases Leanstral 1.5, an updated Lean 4 theorem-proving model

Mistral AI has shipped Leanstral 1.5, an updated version of its specialized model for automated theorem proving and autoformalization in Lean 4, the interactive proof assistant used for mechanically verifying mathematical proofs and program correctness. The model is available now at no cost through Mistral's Labs tier, accessible via the Studio playground and the API.

What's new

According to Mistral's model card, Leanstral 1.5 is described as "an updated Lean 4 formal proof engineering model optimised for automated theorem proving and autoformalization." Key specifications:

• Architecture: 119B total parameters, with 6.5B active — a sparse mixture-of-experts design
• Context window: 256k tokens, long enough to hold an entire proof file and its surrounding code together
• Pricing: free of charge under the Labs v1.5 designation (model ID labs-leanstral-1-5)
• Feature support: chat completions, function calling, agents and conversations, structured/predicted outputs, prefix, OCR and document processing, fill-in-the-middle, embeddings, moderation, audio, and batch processing

The model is positioned specifically for formal verification work rather than general-purpose chat or coding.

Context

Leanstral 1.5 succeeds the original Leanstral, which Mistral introduced in March 2026 as, in the company's words, "the first open-source code agent designed for Lean 4." That release used a highly sparse architecture with 6B active parameters and was framed around a specific engineering bottleneck: as AI coding agents produce more and higher-stakes code, Mistral argued that "the time and specialized expertise required to manually verify become the primary impedance of engineering velocity."

Mistral's stated goal with the Leanstral line is to move code verification away from humans manually debugging machine-generated output and toward agents that can formally prove their own implementations satisfy a specification — in the company's framing, a world where "humans dictate what they want" and the model both writes the code and proves it correct against strict specifications, rather than merely producing text that looks like a valid Lean proof.

Why it matters

Formal verification has historically been a narrow, expert-heavy discipline — Lean 4 proofs are typically hand-built by mathematicians and specialists, not generated at scale. A model purpose-built to interact directly with the Lean 4 compiler, checking and refining its own proof attempts rather than guessing at plausible-looking proof text, points toward verification becoming a more automated part of the software and math-research pipeline rather than a bottleneck on it.

It's also notable as a second consecutive point release in the Leanstral line inside four months, suggesting Mistral sees enough traction in specialized formal-methods models to keep iterating on them alongside its general-purpose lineup — a narrower, more technical bet than its consumer- and enterprise-facing releases, but one aimed at a problem (trust in AI-generated code) that sits upstream of a lot of agentic coding work.