Mistral AI Launches Leanstral 1.5 for Advanced Code Agent Capabilities in Lean 4

TITLE: Mistral AI Launches Leanstral 1.5 for Advanced Code Agent Capabilities in Lean 4
SLUG: mistral-ai-leanstral-1-5-code-agent-lean-4
EXCERPT: Mistral AI has released Leanstral 1.5, an Apache-2.0 licensed code agent model designed for Lean 4, specializing in automated theorem proving and proof engineering. The model demonstrates significant advancements, solving 587 of 672 PutnamBench problems and setting new benchmarks in algebra.
CATEGORY: AI News India
TAGS: Mistral AI, Leanstral, Lean 4, code agent, AI model, theorem proving, Apache 2.0, open-source AI, AI development, India AI
SEO_TITLE: Mistral AI Releases Leanstral 1.5: Advanced Lean 4 Code Agent Model
SEO_DESCRIPTION: Explore Mistral AI’s new Leanstral 1.5, an Apache-2.0 licensed code agent for Lean 4, excelling in automated theorem proving and bug detection, with practical use cases for developers in India.
MEDIA_QUERY: Mistral AI Leanstral 1.5 code agent model for Lean 4, abstract mathematical proofs, AI development environment
IMAGE_ALT: Abstract representation of code and mathematical symbols, illustrating the Leanstral 1.5 AI model’s function in theorem proving and code analysis.

Mistral AI has officially launched Leanstral 1.5, a new code agent model specifically developed for Lean 4, a proof assistant. This release focuses on enhancing automated theorem proving and proof engineering capabilities, making it a significant development for AI researchers and developers. The model is released under an Apache 2.0 license, promoting open access and collaboration, and includes a free API endpoint, `leanstral-1-5`.

Leanstral 1.5 is an update to the previous Leanstral-2603 model and is part of the Mistral Small 4 family. It is designed to mechanically check every logical step in proofs, supporting complex mathematical objects and code properties. The model’s architecture is a Mixture-of-Experts (MoE) with 119 billion parameters, where 6.5 billion parameters are activated per token, ensuring high capacity with efficient computation.

Key Facts

Feature	Detail
Model Name	Leanstral 1.5
Developer	Mistral AI
License	Apache 2.0
Primary Function	Code agent for Lean 4, automated theorem proving, proof engineering
Performance	Solves 587 of 672 PutnamBench problems; saturates miniF2F; new state-of-the-art on FATE-H and FATE-X

Benchmarking and Performance

Leanstral 1.5 demonstrates strong performance across several benchmarks. It achieves 100% on both validation and test sets of miniF2F. Notably, it solves 587 out of 672 problems on PutnamBench, outperforming Seed-Prover 1.5 while being significantly more cost-effective. The model also sets new state-of-the-art results on the FATE-H and FATE-X algebra benchmarks, achieving 87% on FATE-H and 34% on FATE-X. On FLTEval, a benchmark derived from real pull requests to the Fermat’s Last Theorem repository, Pass@1 rises from 21.9 to 28.9, and Pass@8 rises from 31.9 to 43.2, surpassing Opus 4.6 at a fraction of the cost.

Architecture and Training

The model features a context length of 256k tokens and supports multimodal input, accepting both text and images, though its output is text-only. Its training process involves three stages: mid-training, supervised fine-tuning, and reinforcement learning with CISPO. Two distinct reinforcement learning environments were used to shape its agentic behavior. The multiturn environment allows the model to attempt, refine, and re-submit proofs based on Lean compiler feedback. The code agent environment enables Leanstral to interact with a raw filesystem, edit files, execute bash commands, and utilize the Lean language server for real-time goal, error, and type information.

Practical Applications and Use Cases

Beyond theoretical proofs, Leanstral 1.5 has practical applications for engineers. Mistral AI documented two case studies where the model analyzed 57 repositories, flagging 47 violated properties and identifying 11 genuine bugs, five of which were previously unreported on GitHub. One notable bug was found in the sign function for zigzag decoding in `datrs/varinteger`, where an overflow caused crashes and silent data corruption.

These capabilities suggest direct use cases for development teams. Engineers can leverage Leanstral to complete partial proofs within a repository, automatically generate correctness properties for functions, and stress-test Rust code by proving or disproving inferred invariants. This can significantly improve code reliability and reduce debugging time.

Deployment and Accessibility

Mistral AI offers several ways to access Leanstral 1.5. The simplest method is through Mistral Vibe, Mistral’s agent CLI, which allows users to install and launch the Lean agent. For those preferring self-hosting, the model weights can be served using vLLM 0.24.0 or newer, with support for OpenAI-compatible client calls. Users can adjust `reasoning_effort` for complex prompts or faster responses. Leanstral also supports OpenAI-style tool calling, enabling integration with functions like `lean_run_code` for compiling code snippets.

Source: MarkTechPost: “Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems” at https://www.marktechpost.com/2026/07/03/mistral-ai-releases-leanstral-1-5-an-apache-2-0-lean-4-code-agent-model-solving-587-of-672-putnambench-problems/