Lean 4 proof completion
Researchers and formal-methods engineers who want an agent to fill proof holes, build helper lemmas, and iterate against compiler feedback.
Independent tool overview
Leanstral 1.5 at a glance
Leanstral 1.5 is Mistral AI's Apache 2.0 open-weight mixture-of-experts model for Lean 4 formal proof engineering, automated theorem proving, autoformalization, and proof-oriented code agents. It has 119 billion total parameters with 6.5 billion active per token, a 256,000-token context window, downloadable weights, and a free Mistral API preview. Its most important limitation is operational: Mistral says the hosted API model will be retired on September 30, 2026. The weights can remain usable, but self-hosting requires substantial infrastructure and careful validation.
Visit the official Leanstral 1.5 site ↗
- Model type
- Open-weight Lean 4 proof-engineering and theorem-proving model
- Architecture
- Mixture of experts with 119B total parameters and 6.5B active per token
- Experts
- 128 experts, four active per token
- Context window
- 256K tokens; Mistral recommends no more than 200K for the documented self-hosted setup
- Input and output
- Text and image input, text output
- License
- Apache License 2.0
- Hosted model ID
- labs-leanstral-1-5 in current Mistral documentation
- Hosted API price
- Free during the current Labs preview
- Hosted retirement
- September 30, 2026, according to Mistral's changelog
- Recommended agent
- Mistral Vibe with the optional Lean LSP MCP integration
- Reviewed
- August 31, 2026 from Mistral's current model docs, changelog, pricing, release post, model card, and license
Overview
What Leanstral 1.5 is
Leanstral 1.5 is a specialist coding and reasoning model built around Lean 4, not a general-purpose consumer math tutor. It can draft proof terms, repair Lean files, inspect goals and compiler errors, create auxiliary lemmas, navigate repositories, and iterate with a Lean language server until a target proof compiles or its budget is exhausted.
The model is a 119B-parameter mixture of experts with 6.5B parameters active per token, 128 experts and four active experts per token. Mistral documents a 256K context window, recommends keeping self-hosted context at or below 200K, accepts text and image input, and produces text. The model supports high reasoning effort, structured output, function calling, and agent or conversation APIs.
Mistral reports strong results on miniF2F, PutnamBench, FATE-H, FATE-X, and the open-sourced FLTEval proof-engineering benchmark. Those are provider-run results under specific token budgets and tool environments. Some headline outcomes used up to four million tokens per PutnamBench attempt, so benchmark accuracy and practical latency or cost are not interchangeable.
A Lean proof accepted by the kernel is much stronger evidence than a natural-language answer, but it proves only the formal statement in the chosen environment. If the theorem, definitions, assumptions, imported axioms, code translation, or intended specification is wrong, a compiling proof can establish the wrong thing perfectly.
Leanstral 1.5 is available as downloadable Apache 2.0 weights and through Mistral's free Labs API. Mistral's changelog states that the hosted model ID will retire September 30, 2026, only three months after release. New production work should therefore pin weights or a supported deployment path and design migration before depending on the hosted endpoint.
Use cases
Who Leanstral 1.5 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Formal mathematics
Lean users exploring contest, undergraduate, graduate, or research-level formalization with all outputs checked in a pinned environment.
Proof repository maintenance
Teams that need help navigating dependencies, repairing broken proofs, refactoring tactics, or proposing changes across a Lean project.
Code-property verification
Formal-verification teams translating program behavior into Lean specifications and using the model to search for proofs or counterexamples.
Open-weight research
Organizations that need downloadable weights, an Apache 2.0 license, reproducible serving, or experimentation beyond a closed API.
Capabilities
Core Leanstral 1.5 features
Lean 4 specialization
Trained for theorem proving, proof repair, autoformalization, Lean code, and language-server feedback rather than broad conversational math alone.
Compiler-feedback loop
Can submit candidate proofs, read Lean errors and goals, revise the file, and repeat across a multi-turn proof-engineering task.
Repository agent behavior
Can inspect and edit files, run commands, use a Lean language server, add auxiliary lemmas, and persist through long tasks when placed in an authorized agent.
Proof and disproof attempts
Mistral trained the model to try a theorem, react to feedback, and in some workflows attempt a negation or counter-property when proof search fails.
Long context
Provides a documented 256K context window for large proof states and repositories, with 200K or less recommended in the model card's local setup.
High reasoning mode
Supports high reasoning effort for complex tasks and a no-reasoning option for simpler or faster interactions.
Tool calling
Can call configured functions such as a Lean compiler or language-server bridge rather than merely returning untested source text.
Structured outputs
The hosted model supports structured response formats for controlled integration into proof and agent workflows.
Mistral Vibe integration
Mistral provides a Lean agent configuration for its terminal coding agent and recommends the Lean LSP MCP server.
Local vLLM deployment
The official model card documents self-hosting with vLLM, an OpenAI-compatible client, tensor parallelism, and the Mistral reasoning and tool parsers.
Open weights
Weights are downloadable from Hugging Face under Apache 2.0, allowing continued use after the temporary hosted endpoint retires.
Process
How the Leanstral 1.5 workflow works
Step 1
Pin the environment
Record Lean, Mathlib, project commit, dependencies, compiler options, model weights or API ID, serving stack, agent prompt, and tool versions.
Step 2
Write the specification first
Have a qualified mathematician or formal-methods engineer review the theorem, definitions, quantifiers, domains, assumptions, units, and intended connection to reality.
Step 3
Set a narrow target
Identify the exact theorem or proof hole and prohibit unrelated repository changes unless separately approved.
Step 4
Protect the workspace
Use a disposable branch or sandbox, restrict network and secrets, allowlist commands and paths, and never enable automatic approval on valuable systems.
Step 5
Provide local context
Supply relevant imports, neighboring lemmas, style expectations, failed attempts, and the exact Lean goals instead of dumping an entire repository blindly.
Step 6
Connect compiler feedback
Use Lean itself or the Lean LSP MCP so the model receives authoritative type errors and goal states after each proposal.
Step 7
Budget proof search
Set token, time, command, disk, and retry limits; long benchmark-style attempts can consume millions of tokens and many compactions.
Step 8
Compile every result
Accept only code that passes in the pinned clean environment, not a response that says the proof should compile.
Step 9
Audit the proof boundary
Inspect imported axioms, unsafe declarations, admitted results, generated code, trusted translators, and whether the theorem actually expresses the intended claim.
Step 10
Review the diff
Require a human to evaluate readability, maintainability, unnecessary assumptions, proof robustness, licensing, and all changes outside the target.
Step 11
Run independent checks
Rebuild from scratch, run project tests and linters, compare with alternate formulations or counterexamples, and have a domain expert challenge the specification.
Step 12
Plan hosted migration
Before September 30, 2026, test a pinned local deployment or replacement endpoint and document latency, quality, infrastructure, privacy, and rollback.
Cost
Leanstral 1.5 pricing and free plan
Mistral currently lists Leanstral 1.5 input, cached input, and output as free in the Labs API. This is a public-preview price, not a durable production commitment: the hosted model is scheduled for retirement on September 30, 2026. The weights are free to download under Apache 2.0, but local inference, storage, accelerator capacity, engineering, monitoring, and proof-search compute remain the user's cost.
Mistral Labs API
Free
Temporary hosted preview access to Leanstral 1.5.
- Free input tokens
- Free cached input tokens
- Free output tokens
- Mistral account, Labs opt-in, and API key required
- Model is public preview rather than a stable production tier
- Scheduled retirement: September 30, 2026
- Rate limits, availability, and terms should be verified in the account
Open weights
No model-license fee under Apache 2.0
Download and operate Leanstral 1.5 independently.
- 119B total-parameter mixture-of-experts model
- Official model card documents vLLM serving with tensor parallelism
- Compute, accelerator memory, storage, bandwidth, energy, and engineering are separate
- Long proof searches can materially increase operating cost
- Users remain responsible for third-party rights and lawful use
Mistral Vibe
Leanstral can be used from a free Mistral account
Mistral's recommended interactive agent path for Lean proof work.
- Install the Vibe CLI and Lean agent configuration
- Optional Lean LSP MCP connection is recommended
- Do not use the auto-approve or yolo option on sensitive repositories
- Other Vibe features, models, hosted environments, or future plans may have separate terms and limits
Pricing checked . Check current pricing at the source ↗
Assessment
Leanstral 1.5 strengths and limitations
Where it stands out
- Specialized for Lean 4 rather than relying on a general model's incidental theorem-proving ability
- Compiler and language-server feedback can reject malformed proof attempts during the agent loop
- Open weights and Apache 2.0 licensing support inspection, research, customization, and independent hosting
- Free hosted API lowers the cost of evaluating the model before infrastructure work
- Large context supports repository-scale proof engineering and long formal arguments
- Mixture-of-experts design activates only a fraction of total parameters per token
- Tool calling and structured outputs support controlled proof pipelines
- Official Mistral Vibe and Lean LSP MCP guidance provides an end-to-end starting point
- Mistral reports gains across both mathematical benchmarks and proof-engineering repositories
- Formal kernel checking provides materially stronger validation than accepting natural-language mathematical reasoning
- Weights remain available even though the hosted Labs endpoint has a short lifecycle
What to consider
- The hosted Leanstral 1.5 API is scheduled to retire September 30, 2026, making it unsuitable as an unplanned long-term dependency
- A free Labs price is temporary and does not guarantee production rate limits, uptime, support, latency, or future availability
- The downloadable model is 119B total parameters and requires substantial accelerator memory, storage, serving expertise, and operations despite only 6.5B active parameters per token
- Official self-hosting examples use tensor parallelism across four devices, but suitable hardware and throughput still depend on precision, context, concurrency, and serving configuration
- A Lean proof verifies the encoded theorem, not the informal intent, real-world system, original software, or completeness of the specification
- Incorrect definitions, assumptions, quantifiers, abstractions, or code translations can yield a valid proof of the wrong property
- Imported axioms, unsafe features, admitted results, and the trusted computing base must be audited separately
- A model can weaken a statement, add assumptions, redefine a concept, or prove an easier helper theorem while appearing productive
- Failure to find a proof is not evidence that a theorem is false, and failure followed by a claimed negation still requires formal verification
- Provider-reported benchmark results are not independent production guarantees and depend on prompts, tools, token budgets, sampling, and evaluation rules
- Headline PutnamBench scaling used as many as four million tokens per attempt, far beyond ordinary low-latency workflows
- Long context does not guarantee that the model notices every dependency or preserves intent across compaction
- Image input can help with source material but does not guarantee accurate extraction of notation, diagrams, handwriting, or hidden assumptions
- Agent access to a shell, repository, and MCP server creates command execution, secret exposure, destructive-change, dependency, and prompt-injection risks
- Mistral's model card exposes an auto-approve option but explicitly warns users to be careful; production repositories should keep approval and sandbox controls
- Generated proof code can be brittle, opaque, slow to compile, tied to specific library versions, or hard for maintainers to extend
- Formal verification complements rather than replaces tests, fuzzing, static analysis, code review, threat modeling, and runtime monitoring
- The release post's software-bug pipeline depended on Aeneas translation and inferred properties; those extra components expand the trusted and failure boundary
- Local serving may differ from Mistral's API because of quantization, vLLM version, attention backend, precision, parsers, and sampling settings
- Mathematical and safety-critical claims still require qualified expert review even when the final Lean artifact compiles
Compare
Leanstral 1.5 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Project Management
Claude
A general-purpose coding and reasoning assistant with strong repository workflows, but without Leanstral's open weights or dedicated Lean 4 training and kernel-feedback positioning.
Explore Claude →Business Operations
ChatGPT
A broad interactive assistant for explaining mathematics and drafting Lean code when a managed general model is preferable to a specialist open-weight deployment.
Explore ChatGPT →Business Operations
DeepSeek
An open-model ecosystem alternative for mathematical and coding reasoning, though buyers should verify the exact prover model, license, deployment, and Lean-specific benchmark evidence.
Explore DeepSeek →Questions
Leanstral 1.5 FAQs
What is Leanstral 1.5?
Leanstral 1.5 is Mistral AI's open-weight model for Lean 4 formal proof engineering, theorem proving, autoformalization, and proof-oriented coding agents.
Is Leanstral 1.5 open source?
Mistral describes it as open source and publishes the model weights under the Apache License 2.0. The weights and model card are available on Hugging Face.
How large is Leanstral 1.5?
It is a 119B-total-parameter mixture-of-experts model with 6.5B parameters active per token, 128 experts, and four active experts per token.
What is Leanstral 1.5's context window?
Mistral documents a 256K-token context window. The official self-hosted model card recommends using no more than 200K tokens in its sample setup.
Is the Leanstral 1.5 API free?
Yes as of August 31, 2026. Mistral lists input, cached input, and output as free for the Labs model. That preview is scheduled to retire September 30, 2026.
When will Leanstral 1.5 be retired?
Mistral's changelog says the hosted Leanstral 1.5 API model will retire September 30, 2026. The downloadable Apache 2.0 weights are a separate artifact and can continue to be self-hosted.
What model ID should I use?
Current Mistral model documentation lists the hosted ID as labs-leanstral-1-5. Older release copy uses a shorter name, so use the current console and documentation and plan for retirement.
Does a Leanstral proof guarantee the theorem is true?
Lean's kernel can verify that a proof term establishes the formal statement under its definitions and assumptions. It does not guarantee that the statement matches the intended informal claim or real system.
Can Leanstral 1.5 verify software?
It can help prove Lean specifications and Mistral reports experiments on properties translated from Rust. The translation, specification, assumptions, compiler boundary, and connection to deployed code all require independent validation.
Can I run Leanstral 1.5 locally?
Yes. Mistral's model card documents vLLM serving with tensor parallelism and an OpenAI-compatible client. The model is large, so benchmark memory, throughput, context, and total infrastructure cost first.
Does Leanstral replace a mathematician or formal-methods engineer?
No. It can accelerate proof search and Lean editing, but experts must define the right property, review assumptions and the trusted base, inspect changes, and decide whether the proof supports the real-world claim.
Should I use Vibe's auto-approve mode?
Not on sensitive or valuable repositories. Keep the agent sandboxed, restrict tools and paths, require approval for commands and edits, and review every diff before it reaches a trusted branch.
Bottom line
Our Leanstral 1.5 verdict
Leanstral 1.5 is one of the more practically interesting open-weight specialist models because its output can be checked by Lean rather than accepted on style alone. It is best used as a bounded proof engineer: pin the environment, connect authoritative compiler feedback, constrain repository access, and require expert review of the specification and trusted base. The immediate production caveat is decisive—the free hosted endpoint retires September 30, 2026. Evaluate quickly, but depend on it only with a tested self-hosted or replacement path.
Visit Leanstral 1.5 website ↗Similar Science Tools
View all Science AI tools →
Claude Science
Claude Science - Anthropic’s app for AI-powered scientific research
Kosmos
Kosmos - Next-generation AI scientist
TRIBE v2
TRIBE v2 - Meta's predictive foundation model that simulates human brain responses to sights, sounds, and language
Edison Analysis
Edison Analysis - Next-gen scientific analysis agent
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.