Leanstral 1.5

What is Leanstral 1.5?

Leanstral 1.5 is a specialist artificial-intelligence model from Mistral AI focused on Lean 4 formal proof engineering. Lean 4 is a programming language and proof assistant used to express mathematical theorems, software properties, and machine-checkable proofs. Rather than targeting ordinary chat or broad multimodal assistance, Leanstral 1.5 is intended to help users construct, repair, explain, and verify proofs inside real Lean projects.

The model’s canonical API identifier is labs-leanstral-1-5. It belongs to Mistral’s Leanstral family and is positioned as an open-weight model for long-horizon, agentic proof work. In practical terms, that means it can work through a sequence of related tasks: inspect repository files, examine a theorem or unfinished proof, interpret Lean errors, revise code, and continue iterating until the result is accepted by the Lean compiler and proof assistant.

Leanstral 1.5 is available as downloadable weights, through Mistral Vibe, and through a free Labs API endpoint during the documented public-preview period. Mistral released it on June 30, 2026, and announced that it is scheduled for retirement on September 30, 2026. That planned retirement is important for anyone considering a new production integration.

What is Leanstral 1.5 designed to do?

Its primary purpose is formal proof engineering: producing code that is not merely plausible, but checked by Lean 4. The model can assist with theorem proving, proof completion, formalizing mathematical statements, debugging elaboration and compiler errors, and analyzing verified software.

A typical workflow starts with a theorem statement, an incomplete proof, or a repository containing related definitions and lemmas. Leanstral 1.5 can use that context to propose proof steps or Lean code. The surrounding tool environment can then run Lean, return the resulting goals or errors, and give the model information for another attempt. This feedback loop is more important than a single generated answer because a proof is useful only when the proof assistant accepts it.

Mistral describes the model as trained for long-horizon agentic proof engineering, including modifying files, running commands, inspecting Lean goals, and refining proofs over multiple attempts. Mistral recommends using it with Mistral Vibe and, optionally, the Lean language-server MCP server. That setup can expose goals, errors, and type information during development.

Formal reasoning performance

Leanstral 1.5 is specialized for a form of reasoning where the final result can be mechanically checked. This gives it a different practical profile from a general language model that produces an explanation or code sample without verifying the result.

Mistral reports that Leanstral 1.5 achieves 100 percent on the miniF2F validation and test sets, solves 587 of 672 PutnamBench problems under the reported evaluation setup, and scores 87 percent on FATE-H and 34 percent on FATE-X. These are provider-reported benchmark results. They indicate the areas and evaluation conditions Mistral emphasizes, but they are not guarantees that every theorem, repository, or project will be solved successfully.

The model’s strongest reasoning advantage is therefore not simply a high benchmark score. It is the combination of a large context window, specialized training, and iterative access to compiler or proof-assistant feedback. Results can still depend on the quality of the Lean environment, the available libraries, repository organization, theorem difficulty, and how effectively the agent is configured.

Technical specifications

The 256k-token context window is useful for proof repositories because relevant definitions, imports, prior lemmas, error messages, and project instructions can occupy substantial context. The 128k-token maximum output is also unusually large for tasks that may involve long generated files or extended agentic work. These are maximum published limits, not a promise that every request will use or benefit from the full allowance.

Coding, tools, and API support

Leanstral 1.5 supports coding in the specific sense most relevant to formal methods: generating and editing Lean code, proof terms, and related repository content. It is not presented as a general-purpose software-development model, although some of its workflow skills can be useful when formal verification is part of a broader programming project.

Mistral’s official model documentation lists Chat Completions, function calling, Agents and Conversations, and Structured Outputs. Function calling allows an application or agent environment to connect the model to development tools. In a proof workflow, those tools may include repository operations, shell commands, Lean execution, or language-server feedback. Mistral Vibe provides the more complete environment for this style of work.

Structured Outputs can help applications request results in a defined structure. However, the supplied documentation confirms Structured Outputs but does not separately confirm a legacy JSON-mode capability for this exact model. Those two features should not be treated as automatically identical.

Modalities and capability profile

Leanstral 1.5 is a text-in, text-out model. Its input is text such as theorem statements, source files, goals, compiler messages, and instructions. Its output is text that may contain explanations, Lean code, proof terms, or structured text. The supplied research does not verify image, audio, or video input or output for this model.

It is best understood as a reasoning and coding specialist rather than a multimodal assistant. Editorial scoring in the supplied model record rates its reasoning capability highly and its coding capability strongly, while giving it a lower speed score than a lightweight general model and a very high cost score because public-preview API access is free. These are editorial evaluations, not scores published by Mistral and not performance guarantees.

There is no verified model-specific information for streaming, prompt caching, batch API access, user fine-tuning, or a knowledge-cutoff date. Those omissions matter when designing a dependable production service.

Pricing and lifecycle considerations

Leanstral 1.5 is listed as free through the public-preview API endpoint during the documented access period. The research does not establish a recurring paid price for the model, so the free preview should not be interpreted as a permanent commercial rate.

The announced September 30, 2026 retirement date is the most significant practical limitation. Even if the model performs well on a proof repository, an integration created close to that date may require migration or replacement. Users should verify current Mistral lifecycle documentation before committing new production workloads, particularly because public-preview models can change in availability and support.

Best use cases

• Completing or repairing Lean 4 theorem proofs.
• Formalizing mathematical statements into machine-checkable Lean code.
• Investigating Lean compiler, elaboration, and type errors.
• Working through large proof repositories with relevant context spread across many files.
• Verifying algorithms and software properties with a proof assistant.
• Using an agent loop that can inspect files, run commands, receive Lean feedback, and revise its work.
• Research and experimentation with open-weight formal-reasoning models under the Apache 2.0 license.

Limitations and when another option may be better

Leanstral 1.5 is not the right choice for general-purpose conversation, image or audio tasks, or ordinary application coding where formal verification is not required. A broader general-purpose model may be more convenient for everyday writing, unrestricted programming help, or multimodal work. A smaller or faster coding model may also be preferable when the task does not require Lean-specific reasoning or a long repository context.

Even for formal methods, Leanstral 1.5 should not be treated as an autonomous source of guaranteed proofs. Generated proof code must be checked by Lean, and users should review changes for correctness, maintainability, and unintended effects. Performance can vary with project dependencies, imported libraries, available tools, and the quality of feedback supplied to the model.

The short announced lifecycle further favors temporary research, evaluation, and experimentation over long-term dependency on the exact model identifier. If stable production support, a different language ecosystem, or a broader feature set is more important than Lean 4 specialization, another currently supported model may be more appropriate.

When to choose Leanstral 1.5

Choose Leanstral 1.5 when the central problem is Lean 4 proof development and you can provide an environment in which generated work is compiled and checked. Its combination of specialized training, long context, open-weight availability, Apache 2.0 licensing, tool support, and free preview access makes it especially suitable for researchers, formalization projects, proof-assistant experimentation, and repository-level proof engineering.

Choose a different option when you need multimodal output, a general conversational assistant, a guaranteed long-term model endpoint, verified streaming or batch features, or a workflow that does not include formal proof checking. For Lean work, the most important evaluation is not whether the model produces convincing prose; it is whether it consistently produces maintainable proofs that Lean accepts within the time and operational constraints of the project.