Mistral AI has released Leanstral 1.5, a specialized artificial intelligence model designed to automate the conversion of mathematical proofs and software specifications into machine-verifiable formats. The model, released on June 30, 2026, addresses a critical bottleneck in formal verification: the time-consuming process of translating human-readable mathematics into a language that computers can rigorously check.

Why Does Mathematical Proof Verification Matter?

When AI systems generate code or mathematical reasoning, subtle errors can hide beneath a surface appearance of correctness. In high-stakes domains like software verification and mathematical proof, even small logical gaps can cascade into major failures. Formal proof systems like Lean 4 solve this problem by requiring every step of a proof to be written in a way that computers can mechanically verify, leaving no room for hidden assumptions or logical leaps.

However, this rigor comes at a cost. Mathematicians and engineers naturally write proofs using shorthand expressions like "obviously true" or "can be proven the same way." Lean 4 cannot understand these human conventions. Converting a research paper or software specification into definitions and theorems that Lean 4 accepts is a painstaking, specialized task that has historically limited formal verification to experts with deep technical training.

What Makes Leanstral 1.5 Different From Its Predecessor?

Leanstral 1.5 is the successor to an earlier version released in March 2026, which has now been discontinued. The new model employs a mixed expert (MoE) architecture, a design pattern that activates only the most relevant portions of a neural network for each task. Leanstral 1.5 contains 119 billion total parameters, but only 6.5 billion parameters are active during processing, making it more efficient than a traditional model of the same size.

The model also supports a context length of 256,000 tokens, roughly equivalent to processing 200,000 words at once. This extended memory allows developers to work with long proof files and related code together, reducing the need to break large verification tasks into smaller chunks.

How to Use Leanstral 1.5 for Automated Theorem Proving and Formalization

• Automated Theorem Proving: A developer inputs the goal of a mathematical proof, and Leanstral 1.5 proposes the necessary proof steps, which Lean 4 then verifies for logical correctness. This workflow reduces the manual effort required to construct formal proofs from scratch.
• Automated Formalization: The model converts human-readable mathematical explanations and specifications from research papers or software documentation into definitions and theorems that Lean 4 can understand, eliminating the need for manual rewriting.
• API Integration: Leanstral 1.5 supports standard API features including Chat Completions, Function Calling, and Agents and Conversations, allowing developers to integrate the model into existing workflows and automation pipelines.

By reducing the burden of converting natural language mathematics into formal notation, Leanstral 1.5 has the potential to extend formal verification beyond a small group of specialists to broader practical verification work in industry and research.

How Is Leanstral 1.5 Available to Users?

Mistral AI is offering Leanstral 1.5 for free through its Labs tier, making it accessible to researchers, developers, and organizations interested in formal verification without requiring a paid subscription. Users can test the model directly in the Mistral AI playground.

However, the release announcement does not include new performance benchmarks comparing Leanstral 1.5 to its March predecessor, nor has Mistral published comparison results against other formal verification tools or policies regarding the release of model weights. This lack of detailed performance metrics makes it difficult for users to assess exactly how much the model has improved or how it compares to alternative approaches.

The release of Leanstral 1.5 reflects a broader trend in AI development: specialized models tailored to specific, high-value tasks. Rather than building one general-purpose model, companies like Mistral are creating focused tools that excel in narrow domains where accuracy and reliability are non-negotiable. For mathematical research, software engineering, and formal verification, this targeted approach may prove more valuable than a generalist alternative.