Mistral AI Unleashes Leanstral 1.5: A Game-Changer for Automated Proof Engineering

Mistral AI‘s Leanstral 1.5: Revolutionizing Automated Proof Engineering

At a glance, In a significant stride for artificial intelligence in software development, Mistral AI has unveiled Leanstral 1.5. This advanced code agent model, designed specifically for the Lean 4 proof assistant, marks a new era in automated theorem proving and proof engineering. Released under an open Apache 2.0 license, Leanstral 1.5 not only pushes the boundaries of what AI can achieve in formal verification but also offers a free API endpoint for developers.

Meanwhile, Leanstral 1.5 builds upon its predecessor, Leanstral-2603, and is an integral part of the powerful Mistral Small 4 family of models. Its release signifies Mistral AI’s commitment to making sophisticated AI tools accessible and impactful for the developer community.

Understanding Leanstral 1.5: A Deep Dive

At its core, Leanstral 1.5 is a sophisticated code agent model tailored for Lean 4. For those unfamiliar, Lean 4 is a proof assistant—a tool that meticulously checks every logical step in a proof, ensuring mechanical correctness. This capability allows Lean 4 to express complex mathematical objects, such as perfectoid spaces, and verify properties of programming language fragments like those found in Rust.

The Mixture-of-Experts Architecture

In practical terms, What powers Leanstral 1.5’s impressive capabilities is its innovative Mixture-of-Experts (MoE) architecture. This design intelligently routes each input token to a select few specialized sub-networks, or ‘experts.’ This approach significantly reduces computational overhead while maintaining a vast total capacity. Leanstral 1.5 employs 128 experts, with only 4 actively engaged per token, optimizing efficiency.

• Total Parameters: 119 Billion
• Active Parameters per Token: 6.5 Billion
• Context Length: 256,000 tokens
• Input Modality: Multimodal (text and image)
• Output Modality: Text only

The Rigorous Training Journey of Leanstral 1.5

Mistral AI developed Leanstral 1.5 through a meticulous three-stage training process, designed to imbue the model with robust agentic behavior:

1. Mid-training: Initial foundational learning.
2. Supervised Fine-Tuning (SFT): Refining the model’s understanding with expert-labeled data.
3. Reinforcement Learning with CISPO: Further enhancing its problem-solving abilities.

For example, Two distinct reinforcement learning environments were crucial in shaping Leanstral’s agentic prowess:

• Multiturn Environment: Here, the model receives a theorem statement and must prove or disprove it. It submits a proof, analyzes feedback from the Lean compiler, and iteratively refines its attempts until successful or its budget is exhausted.
• Code Agent Environment: In this more complex setting, Leanstral operates within a raw filesystem. It can edit files, execute bash commands, and leverage the Lean language server to access real-time goals, errors, and type information. This environment allows it to complete partial proofs, construct auxiliary lemmas, and manage context effectively even during long, complex tasks through context compaction.

Mistral’s custom fork of SafeVerify is used to rigorously verify the correctness of Leanstral’s outputs against target theorems.

Unrivaled Performance and Benchmarks

That said, Leanstral 1.5 demonstrates exceptional performance across several challenging benchmarks, setting new state-of-the-art records:

• miniF2F: Achieves a perfect 100% on both validation and test sets.
• PutnamBench: Solves an impressive 587 out of 672 problems, showcasing its advanced mathematical reasoning.
• FATE-H & FATE-X: Establishes new state-of-the-art results with 87% on FATE-H and 34% on FATE-X, both algebra benchmarks.
• FLTEval: Significantly improves pass@1 from 21.9 to 28.9 and pass@8 from 31.9 to 43.2. FLTEval problems are derived from real pull requests to the Fermat’s Last Theorem repository.

Notably, on FLTEval, Leanstral 1.5 surpasses Opus 4.6’s 39.6 at one-seventh of the cost, and it outperforms open-source models three to ten times larger by a significant margin. Its efficiency is also evident on PutnamBench, where it solves problems at approximately $4 each, a stark contrast to competitors like Seed-Prover 1.5, which can cost upwards of $300 per problem.

Interestingly, The model’s ability to scale its performance with increased token budget is a defining characteristic, with more complex problems being solved as computational resources are expanded per attempt.

Real-World Impact and Practical Use Cases

While primarily trained on mathematics, Leanstral 1.5’s capabilities extend to verifying code, offering tangible benefits for engineers.

Case Study 1: AVL Tree Complexity Proof

However, Leanstral successfully proved the O(log n) time complexity for a real-world AVL tree implementation. This complex task involved structural induction and monadic time tracking, processing over 2.7 million tokens across 22 context compactions to establish a precise bound.

Case Study 2: Bug Detection in Open-Source Code

Perhaps most compelling for developers, Leanstral 1.5 demonstrated its ability to find genuine bugs in open-source code. An automated pipeline used Aeneas to translate Rust code into Lean, after which Leanstral inferred user intent and generated correctness properties.

By attempting to prove or disprove these properties, the model flagged 47 violated properties and identified 11 genuine bugs across 57 repositories, five of which were previously unreported on GitHub. One critical bug involved an overflow in a zigzag decoding function for datrs/varinteger, leading to crashes or silent data corruption.

Meanwhile, “Leanstral 1.5’s ability to find previously unreported bugs in real-world code highlights its profound potential to enhance software reliability and security.”

These examples translate directly into practical use cases for development teams:

• Completing partial proofs within a codebase.
• Automatically generating correctness properties for functions.
• Stress-testing Rust code by verifying or refuting inferred invariants.

Getting Started with Leanstral 1.5

In practical terms, Accessing Leanstral 1.5 is straightforward, catering to both quick experimentation and deeper integration:

Mistral Vibe (Agent CLI)

The easiest way to get started is through Mistral Vibe, Mistral’s agent command-line interface. Leanstral runs on Mistral’s free plan. Simply enable ‘Labs models’ in your account, create an API key, and then use the following commands:

uv tool install mistral-vibe
uv tool update mistral-vibe
vibe –setup

/leanstall
exit

vibe –agent lean

Self-Hosting with vLLM

For example, For those requiring more control or scale, Leanstral 1.5 weights can be self-hosted using vLLM (version 0.24.0 or newer). After installation, serve the weights:

uv pip install -U vllm –torch-backend=auto

vllm serve mistralai/Leanstral-1.5-119B-A6B
–max-model-len 200000
–tensor-parallel-size 4
–attention-backend FLASH_ATTN_MLA
–tool-call-parser mistral
–enable-auto-tool-choice
–reasoning-parser mistral

You can then interact with the server using an OpenAI-compatible client, adjusting the reasoning_effort for speed or complexity:

from openai import OpenAI

client = OpenAI(api_key=”EMPTY”, base_url=”<your-host-url>”)

TEMP = 1.0
MAX_TOK = 32000
REASONING = “high” # switch to ‘none’ for faster answers

model = client.models.list().data[0].id

messages = [
{“role”: “user”, “content”: [
{“type”: “text”, “text”: “Define the transition rules as an inductive proposition in Lean 4.”}
]},
]

response = client.chat.completions.create(
model=model,
messages=messages,
temperature=TEMP,
max_tokens=MAX_TOK,
reasoning_effort=REASONING,
)

print(response.choices[0].message.content)
print(response.choices[0].message.reasoning)

That said, Leanstral also supports OpenAI-style tool calling, allowing developers to integrate functions like lean_run_code for compiling snippets. For tighter Lean integration, the lean-lsp-mcp server is recommended.

Expert Perspective

A practical read on Leanstral 1.5 starts with leanstral. That is where the earliest effects are likely to show up if this development keeps building.

What happens next will come down to adoption speed, policy response, and execution quality. That combination could make Leanstral 1.5 a meaningful reference point across proof.

For decision-makers, the useful lens is not the headline alone but how model changes priorities once organizations have to respond.

Frequently Asked Questions

Why is Leanstral 1.5 important?

Mistral AI’s Leanstral 1.5: Revolutionizing Automated Proof EngineeringAt a glance, In a significant stride for artificial intelligence in software development, Mistral AI has unveiled Leanstral 1.5.

What impact could Leanstral 1.5 have?

This advanced code agent model, designed specifically for the Lean 4 proof assistant, marks a new era in automated theorem proving and proof engineering.

What should readers watch next with Leanstral 1.5?

Released under an open Apache 2.0 license, Leanstral 1.5 not only pushes the boundaries of what AI can achieve in formal verification but also offers a free API endpoint for developers.Meanwhile, Leanstral 1.5 builds upon its predecessor, Leanstral-2603, and is an integral part of the powerful Mistral Small 4 family of models.

How does this relate to leanstral?

It connects because the article frames leanstral as one of the clearest areas where the topic may be felt in practice.

Key Takeaways

What matters next is how the immediate response turns into lasting change. Mistral AI’s Leanstral 1.5 represents a monumental leap in AI for formal verification and software engineering. Here are the core highlights:

• Leanstral 1.5 is an open, Apache-2.0 licensed Lean 4 proof-engineering model.
• It leverages a 119B Mixture-of-Experts architecture with 6.5B active parameters per token.
• The model achieves 100% saturation on miniF2F and solves 587 of 672 PutnamBench problems.
• It has a proven track record of identifying real-world bugs, finding 5 previously unreported issues in open-source repositories.
• Developers can access it via a free API endpoint (Mistral Vibe), self-host its weights via vLLM, or utilize its Hugging Face presence.

Interestingly, Leanstral 1.5 is poised to become an indispensable tool for researchers and developers aiming to achieve unprecedented levels of correctness and reliability in their software and mathematical proofs.

Source: 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/