Mistral Releases Leanstral 1.5: Pushing the Boundaries of Formal Mathematical Proofs in Europe

Bridging Neural Intuition and Mathematical Rigor

While general-purpose frontier models excel at creative writing, code generation, and informal problem-solving, they frequently struggle when forced to adhere strictly to formal logic. In informal mathematics, a plausible-sounding explanation can obscure subtle errors. In formal mathematics, however, every statement must be verified down to fundamental axioms by an interactive theorem prover (ITP) such as Lean 4, Coq, or Isabelle.

With Leanstral 1.5, Mistral AI builds on its heritage of domain-focused models (such as Codestral and Mathstral) to deliver a tool designed expressly for formal logic synthesis. The model acts as an intelligent proof copilot, generating tactic scripts that are immediately checked against the Lean 4 environment REPL. When the theorem prover rejects a step, Leanstral 1.5 consumes the precise feedback error state to self-correct its proof trajectory.

What Is "Proof Abundance" and Why Does It Matter?

The central technical innovation in Leanstral 1.5 is its architectural emphasis on Proof Abundance. Previous iterations of automated theorem provers relied on narrow beam searches or low-throughput sampling, which often stalled when encountering complex lemmas. Leanstral 1.5 fundamentally alters this dynamic by combining high-throughput parallel sampling with deep tactical search trees.

Instead of aiming for a single canonical proof on the first attempt, the model generates dozens or hundreds of candidate proof branches in parallel. By coupling this generation with automated Lean 4 verification, Leanstral 1.5 rapidly filters out invalid tactics and amplifies correct logical branches. This approach yields two major advantages:

• Higher Solve Rates: Complex mathematical benchmarks that previously required human mathematician intervention can now be solved through sheer computational exploration of verified formal paths.
• Synthetic Data Generation: The model produces thousands of verified formal proof scripts, creating a self-reinforcing flywheel for training future open-weights reasoning models across Europe.

Benchmark Comparison: Formal Verification Performance

To evaluate Leanstral 1.5's capabilities, performance on key formal mathematical benchmarks—most notably miniF2F (a benchmark containing formal Olympiad and high-school competition problems in Lean 4) and PutnamBench—provides a clear picture of how it compares against existing models.

The benchmark data highlights the decisive advantage of specialized formal training: while massive generalist models like GPT-4o achieve under 30% on direct single-pass tactic generation, Leanstral 1.5 leverages its proof abundance architecture to cross 67% under multi-sample pass strategies.

European Industrial Relevance: The EU AI Act and Deterministic Code

For businesses operating within the European Union, the launch of Leanstral 1.5 carries implications that extend far beyond pure academic mathematics. Under the framework of the EU AI Act, safety-critical software—such as medical device firmware, automotive control modules (AUTOSAR), aerospace software, and financial smart contracts—is subject to stringent standards regarding risk mitigation, transparency, and verifiable safety.

Traditional generative code models produce probabilistic code that may contain subtle, dangerous edge-case bugs. By combining generative AI with Lean 4 formal verification, European engineering teams can automatically write code alongside mathematical proofs confirming that the software strictly satisfies its specification guarantees.

Furthermore, because Leanstral 1.5 is released under an open-weights model, European enterprises can deploy it entirely on-premise within local sovereign cloud infrastructure, ensuring full compliance with European Data Protection Regulations (GDPR) and preserving corporate intellectual property.

Availability and API Pricing (EUR & USD)

Leanstral 1.5 is immediately available through two primary channels across all 27 EU member states and globally:

1. Open Weights Download: The model weights can be downloaded directly via Hugging Face for local hosting on enterprise GPU clusters or developer workstations.

2. Mistral La Plateforme Cloud: Hosted on European infrastructure hosted in France and Germany, offering full European data residency guarantees.

Pricing for the La Plateforme cloud endpoint is structured competitively for enterprise high-throughput verification pipelines:

• Input Tokens: €0.00015 / $0.00016 per 1,000 tokens
• Output Tokens: €0.00045 / $0.00048 per 1,000 tokens

Compared to closed proprietary models, this pricing structure allows developer teams to run high-volume sampling routines (generating hundreds of proof variants per problem) at a fraction of standard API costs.

Deploying Leanstral 1.5 Locally

For developers interested in setting up a local formal verification environment, Leanstral 1.5 can be integrated with local inference backends such as vLLM or Ollama paired with the Lean 4 REPL server. Developers running hardware setups—such as local workstations equipped with NVIDIA RTX 50 series GPUs—can run high-throughput parallel tactic searches entirely offline without transmitting sensitive codebase specifications to cloud servers.