Leanstral - Mistral AI's first open-source AI code agent

What is Leanstral

Leanstral is Mistral AI's first open-source AI code agent, designed specifically for the Lean 4 theorem prover. The model uses a sparse architecture with 120B total parameters / 6B activated parameters, and can automatically generate formal proofs and verify code correctness. Compared with giant competitors, Leanstral achieves higher efficiency at extremely low cost ($18 per run), and performs excellently in evaluations on real mathematical codebases (such as the Fermat's Last Theorem project). The model supports extension through the MCP protocol and has been integrated into the Mistral Vibe platform.

Leanstral's main features

Automatic formal proof generation: For the Lean 4 proof assistant, automatically generates rigorous mathematical proofs and software specification verification code.
Code correctness verification: Through Lean 4's complete verifier, ensures that generated code conforms to strict formal specifications, eliminating the manual review bottleneck.
Intelligent diagnosis and repair: Supports analyzing the reasons for code failures (such as identifyingdefandabbrevdifferences in type aliases), and provides precise repair solutions.
Cross-language conversion: Supports automatically converting other proof languages (such as Rocq/Coq) into Lean 4 code, preserving custom notation.
Theorem proving assistance: Completes formal proofs and definitions of new mathematical concepts in real mathematical codebases (such as the Fermat's Last Theorem project).

Leanstral's key information and usage requirements

Developer：Mistral AI
Positioning: The first open-source AI code agent designed specifically for Lean 4
Architecture: Sparse mixture of experts (MoE), 120B total parameters / 6B activated parameters
License: Apache 2.0 (fully open source)
Cost: $18 per run, only $36 for pass@2 (compared with Claude Sonnet at $549)
Performance: FLTEval score 29.3 (pass@4), surpassing most open-source competitors
Mistral Vibe: Zero-configuration integration, usable by entering /leanstall
Labs API: Free/low-cost endpoint labs-leanstral-2603 (available for a limited time)
Local deployment: Download the open-source weights and run them yourself

Leanstral's core advantages

Extreme efficiency: Surpasses hundred-billion-level open-source models with only 6B activated parameters, achieving the best balance between performance and cost.
Cost revolution: Requires only $18 per task, achieving better verification results at 1/15 the price of Claude Sonnet.
Fully open source: Opens weights under the Apache 2.0 license, eliminating vendor lock-in and supporting private deployment and autonomous control.
Vertical optimization: Deeply trained for Lean 4 proof engineering, performing far beyond general large models in real mathematical codebases.
Trustworthy verification: Supports generated code accompanied by formal mathematical proofs, transforming the manual review bottleneck into automatic machine verification.
Ecosystem compatibility: Natively supports the MCP protocol and can seamlessly integrate with existing development toolchains and language servers.

How to use Leanstral

Mistral Vibe (recommended for beginners): Visit the Mistral Vibe platform and enter in the conversation /leanstall command to start with zero configuration, no need to install any local environment.
Labs API (Developer): Call the API endpoint labs-leanstral-2603, currently free for a limited time, suitable for integration into automated workflows or custom applications.
Local Deployment (Advanced Users): Download the Apache 2.0 licensed model weights from official channels and run them independently on your own hardware, achieving complete data privacy and control.
Usage Recommendations: Best performance is achieved with the lean-lsp-mcp tool, suitable for scenarios such as formal mathematical proofs and high-assurance software verification.

Leanstral's project address

Project official website：https://mistral.ai/news/leanstral

Comparison of Leanstral with similar competitors

Comparison dimension	Model	Scale	FLTEval score	Cost	Features
Leanstral	Leanstral-120B-A6B	120B/6B	26.3 (pass@2) 29.3 (pass@4) 31.9 (pass@16)	$18-$290	Optimized specifically for Lean 4, open source, MCP extension
Open-source competitor	Qwen3.5-397B-A17B	397B/17B	25.4 (pass@4)	–	Requires 4 rounds to achieve the effect of Leanstral in 2 rounds
	Kimi-K2.5-1T-A32B	1T/32B	20.1 (pass@4)	–	Huge scale but obvious score bottleneck
	GLM5-744B-A40B	744B/40B	16.6 (pass@4)	–	Largest parameter count but worst performance
Closed-source competitor	Claude Opus 4.6	–	39.6	$1,650	Highest quality but cost is 92 times that of Leanstral
	Claude Sonnet 4.6	–	23.7	$549	Cost is 15 times that of Leanstral, with lower scores
	Claude Haiku 4.5	–	23.0	$184	Mediocre cost-effectiveness

Leanstral's application scenarios

Formal mathematical proofs: Automatically complete formal proofs in large mathematical projects such as Fermat's Last Theorem, and correctly define new mathematical concepts.
High-assurance software verification: Verify strict properties of code snippets in programming languages such as Rust, ensuring software correctness for mission-critical systems.
Codebase migration and adaptation: Diagnose and fix breaking changes caused by Lean version upgrades, for example automatically identifyingdefandabbrevdifferences in type aliases and providing fixes.
Cross-language code conversion: Fully convert code in other proof languages such as Rocq/Coq to Lean 4, preserving custom notation and logical structure.
Intelligent debugging and diagnosis: The model supports analyzing the root causes of compilation failures, automatically generating test cases to reproduce problems, and providing precise fixes and explanations of the underlying principles.