—
**Article: “Mistral AI Introduces Leanstral 1.5—A Code Agent Model for Lean 4 Theorem Proving”**
By Marktechpost
Published: [Date Redacted] | Last updated: [Date Redacted]
*(Cross-posted with permission from Marktechpost)*
—
In a recent development for formal methods and automated reasoning, Mistral AI has announced the release of **Leanstral 1.5**, a specialized code agent model designed for **Lean 4**, a leading proof assistant. The model focuses on **automated theorem proving (ATP)** and **proof engineering**, with open weights under the **Apache 2.0** license and a free API endpoint (`leanstral-1-5`) now available.
### What is Leanstral 1.5?
Leanstral 1.5 is a **Mixture-of-Experts (MoE)** model tailored for Lean 4, a proof assistant capable of expressing sophisticated mathematical objects such as perfectoid spaces and even properties of Rust fragments. The model architecture uses **128 experts**, with **4 active per token**, resulting in **119 billion total parameters** and **6.5 billion activated per token**. It supports a **256k token context length**, is **multimodal** (accepts text and image input), and generates text-only output.
### How Mistral Trained Leanstral 1.5
Training follows a three-stage pipeline:
1. **Mid-training**
2. **Supervised Fine-Tuning (SFT)**
3. **Reinforcement Learning with CISPO**
Two reinforcement-learning environments shaped the model’s agentic behavior:
– **Multiturn environment**: The model receives a theorem, attempts to prove or disprove it, and iteratively refines based on Lean compiler feedback until success or budget exhaustion.
– **Code agent environment**: Leanstral operates in a raw filesystem, editing files, running bash commands, and interacting with the Lean language server to complete partial proofs, build auxiliary lemmas, and persist through context compression.
### Benchmark Performance
Leanstral 1.5 demonstrates strong empirical results:
– **miniF2F**: 100% on both validation and test sets (saturated).
– **PutnamBench**: Solves **587 out of 672** competition problems at approximately **$4 per problem**.
– **FATE-H**: **87%**, a new state-of-the-art for graduate-level abstract algebra.
– **FATE-X**: **34%**, a new state-of-the-art for PhD-level abstract algebra.
– **FLTEval**:
– Pass@1: **28.9%** (up from 21.9%)
– Pass@8: **43.2%** (up from 31.9%), surpassing **Opus 4.6** (39.6%) at one-seventh the cost.
On PutnamBench, Leanstral 1.5 outperforms **Seed-Prover 1.5 high** by 7 problems at a fraction of the cost (estimated at ~$300+ per problem for the latter). Compared to other systems such as **Goedel-Architect** and **AxProverBase**, and with **Aleph Prover** costing roughly $54–$68 per problem, Leanstral 1.5 offers a cost-effective high-performance alternative.
### Test-Time Scaling
A notable feature of Leanstral 1.5 is **test-time scaling**: increasing the token budget per attempt consistently improves performance. For example:
– 50k tokens → 44 solved
– 200k tokens → 244 solved
– 1M tokens → 493 solved
– 4M tokens → 587 solved (full set)
### Interactive Exploration
The release includes an **interactive explorer** that visualizes the scaling curve, benchmark results, cost breakdown, and a simulation of the multiturn proof verification loop. The widget is embeddable and runs in-browser.
### Availability
– Model weights: **Apache 2.0**
– API endpoint: `leanstral-1-5` (free)
– Open-source ecosystem support via Hugging Face and official documentation at **docs.mistral.ai**
—
### Source
> Mistral AI. “Leanstral 1.5 — A Leap in Automated Theorem Proving.” *mistral.ai/news/leanstral-1-5*, updated [date]. https://docs.mistral.ai/news/leanstral-1-5
> *(This article is based on the official Mistral AI announcement and interactive demo documentation.)*
—
