Leanstral Review: Mistral's Open-Source Lean 4 Proof Agent
Mistral releases the first open-source AI agent purpose-built for Lean 4 formal verification. Leanstral is a 120B sparse MoE model (6B active parameters) with an Apache 2.0 license, free API, and a 15× cost advantage over Claude Sonnet 4.6 on real-world proof engineering benchmarks.
Reviewed 2026-06-14 · Source: mistral.ai/news/leanstral
Key Specs
What is Leanstral?
Leanstral is Mistral AI's open-source code agent for Lean 4 — the interactive proof assistant used to formally verify mathematical proofs and software specifications. It was released March 16, 2026, and is the first model specifically trained for proof engineering in realistic Lean 4 repositories.
The model uses a sparse Mixture-of-Experts architecture: 120B total parameters with only 6B active per forward pass. This makes it highly efficient — it outperforms models many times its active size on formal verification tasks while running at a fraction of the cost. A pass@2 run costs $36, compared to $549 for a single Claude Sonnet 4.6 run on the same benchmark.
Unlike general coding models that can generate Lean 4 syntax, Leanstral was trained on formal proof repositories and evaluated on real-world proof engineering tasks — not isolated competition math. The benchmark Mistral built, FLTEval, measures performance on completing proofs in PRs to the FLT project (a Lean 4 formalization of Fermat's Last Theorem), which requires long-horizon reasoning and repository context.
What's notable
Purpose-built for Lean 4 proof engineering
Leanstral isn't a general coding model with Lean 4 bolted on — it's trained specifically for the Lean 4 proof assistant and formal verification workflows. This matters because Lean 4 has its own type system, tactic language, and definitional equality rules that differ fundamentally from general-purpose code. Existing models that 'support' Lean 4 are generalist models that happen to generate Lean syntax. Leanstral was trained on realistic formal repositories — including the FLT (Fermat's Last Theorem) formalization project — meaning it understands proof engineering at scale, not just isolated math competition problems.
Extreme cost efficiency via sparse MoE architecture
Leanstral uses a sparse Mixture-of-Experts architecture: 120B total parameters, but only 6B are active on any given inference pass. This gives it a quality profile approaching much larger dense models while running at a fraction of the compute cost. On Mistral's FLTEval benchmark, Leanstral at pass@2 scores 26.3 — beating Claude Sonnet 4.6's 23.7 at a total cost of $36 vs Sonnet's $549. That's a 15× cost advantage for better performance on the task that matters. At pass@16, Leanstral reaches 31.9 compared to Sonnet's 23.7 — an 8-point advantage while still costing less than one-fifth of Sonnet's single-pass cost.
Apache 2.0 open weights with free API endpoint
Mistral released Leanstral weights under the Apache 2.0 license — no usage restrictions, no commercial licensing, no fine-tuning prohibitions. For research labs, universities, and companies building formal verification pipelines, this means full control over the model stack. Mistral also provides a free API endpoint at launch (via La Plateforme), so teams don't need to spin up their own inference infrastructure to get started. Leanstral is also integrated directly into Mistral Vibe — activate it with /leanstall for zero-configuration proving in the agentic coding environment.
MCP-native design with lean-lsp-mcp support
Leanstral supports arbitrary MCPs (Model Context Protocol tools) through Mistral Vibe, and was specifically trained to achieve peak performance with the lean-lsp-mcp — the language server protocol for Lean 4. This means Leanstral can call the Lean compiler inline, check proof states, and iterate on tactics with the same feedback loop a human developer would use. MCP-native design also means teams can add custom tools (specification checkers, domain-specific evaluators) without retraining the model.
New FLTEval benchmark for real proof engineering
Existing Lean 4 benchmarks focus on competition math — isolated, self-contained problems like miniF2F. Mistral built FLTEval to measure something more realistic: completing all formal proofs and correctly defining new mathematical concepts in each PR to the FLT (Fermat's Last Theorem formalization) project. This is a long-horizon, repository-scale task where context and proof engineering judgment matter more than raw symbol manipulation. Mistral is releasing FLTEval publicly alongside Leanstral, giving the broader research community a better signal for comparing formal verification agents.
FLTEval Benchmark: Score vs. Cost
FLTEval measures proof completion on real PRs to the Lean 4 FLT project. Higher score is better. Cost is per benchmark run (not per token).
Source: Mistral AI, March 2026. Leanstral rows highlighted. OSS model costs not published by Mistral.
Pricing & Access
Available via Mistral La Plateforme. Check mistral.ai/pricing for current rate details. Same API key as other Mistral models.
Activate with /leanstall in Mistral Vibe. No infrastructure setup needed. Best for individual researchers and developers getting started.
Download weights from HuggingFace. Apache 2.0 — commercial use permitted. Suitable for teams with on-prem inference requirements.
Who is Leanstral for?
Leanstral is a highly specialized model. If you don't work with Lean 4, it's not relevant to your workflow. But for the formal verification community, it's a significant release.
The primary use cases are: (1) Mathematicians and researchers working on large Lean 4 proof repositories who need to automate proof completion; (2) Software engineering teams applying formal verification to mission-critical code — Lean 4 can express and verify properties of Rust fragments and other safety-critical software; (3) Universities and research labs that need an open-weight model for formal methods coursework or research without API cost constraints.
For general coding tasks, Devstral 2 or Codestral 25.08 are more appropriate. For reasoning or general LLM use, Mistral's other models (Mistral Medium 3.5, Mistral Small 4) serve those needs. Leanstral is for Lean 4, specifically.
FAQ
What is Leanstral?
Leanstral is Mistral AI's open-source AI agent purpose-built for Lean 4, the interactive proof assistant used for formal verification of mathematics and software. Released March 16, 2026, it's a 120B-parameter sparse MoE model with 6B active parameters. It was designed to work in realistic formal repositories — not just isolated math competition problems — and is available under an Apache 2.0 license with a free API endpoint.
How does Leanstral compare to Claude on Lean 4 tasks?
On Mistral's FLTEval benchmark (PR-level formal proof completion in the Fermat's Last Theorem formalization project), Leanstral at pass@2 scores 26.3 vs Claude Sonnet 4.6's 23.7 — while costing $36 vs $549 (15× cheaper). At pass@16, Leanstral reaches 31.9 vs Sonnet's 23.7. Claude Opus 4.6 still leads at 39.6, but at $1,650 per benchmark run vs $290 for Leanstral pass@16.
Is Leanstral open source?
Yes. Leanstral weights are released under the Apache 2.0 license, which permits unrestricted commercial use, modification, and distribution. The weights are available for self-hosting and a tech report explaining the training methodology is also being released. There is also a free API endpoint via Mistral La Plateforme for teams that don't need to self-host.
What is FLTEval?
FLTEval is a new evaluation benchmark developed by Mistral to measure formal verification agents on realistic proof engineering tasks. Rather than using isolated competition math problems (like miniF2F), FLTEval tasks models with completing all formal proofs and correctly defining new mathematical concepts in each PR to the FLT project — Lean 4's formalization of Fermat's Last Theorem. Mistral is releasing FLTEval publicly alongside Leanstral.
What is Lean 4?
Lean 4 is an interactive proof assistant and functional programming language developed at Microsoft Research and Lean FRO. It lets developers write formal mathematical proofs and formally verify software against strict specifications. It's used in academic mathematics (including the formalization of complex results like Fermat's Last Theorem) and in software engineering for mission-critical correctness requirements.
How do I use Leanstral?
There are three ways: (1) Mistral Vibe — activate with /leanstall for zero-setup proving in the agentic coding environment; (2) Mistral API — available via La Plateforme with a free endpoint at launch; (3) Self-hosting — download the Apache 2.0 weights from HuggingFace and run your own inference. Leanstral also supports lean-lsp-mcp for full language server integration.
Who is Leanstral for?
Leanstral is for mathematicians, formal methods researchers, and software engineers working in Lean 4. Practical use cases include: automated proof completion in large Lean repositories, formal verification of Rust or other safety-critical software specifications, academic research on theorem proving, and teams exploring AI-assisted formal methods. It is not a general-purpose coding model.
Try Leanstral
Free API endpoint and open weights under Apache 2.0. Available in Mistral Vibe with zero setup.
View Leanstral on MistralAffiliate disclosure: Some links on this page are affiliate links. If you sign up through them, AISO Tools may earn a commission at no extra cost to you. This never affects our rankings or reviews.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.
Join thousands of professionals who discover the best AI tools every week. No spam — unsubscribe anytime.