Leanstral: Open-source agent for trustworthy coding and formal proof engineering
TL;DR Highlight
Mistral released Leanstral, an Apache 2.0 AI agent for Lean 4 mathematical formal proofs, achieving comparable or better performance than Claude Sonnet at 1/15th the cost.
Who Should Read
Developers and researchers applying AI code generation to mission-critical software or mathematical research where accuracy verification is a bottleneck. Researchers in math/CS interested in Lean 4 or formal verification.
Core Mechanics
- Human review of AI-generated code is the biggest engineering speed bottleneck. Leanstral is an agent that solves this by proving mathematically rigorous specifications alongside code generation.
- Lean 4 is a proof assistant that can express both complex mathematical structures (like perfectoid spaces) and software specifications (like Rust code properties). Leanstral is the first open-source code agent designed specifically for Lean 4.
- MoE (Mixture of Experts) architecture with 120B total parameters but only 6B activated during inference for high efficiency. Weights are fully released under Apache 2.0 license.
- Published FLTEval, a new benchmark evaluating real Fermat's Last Theorem project PRs — completing formal proofs and correctly defining new mathematical concepts, moving beyond competition math problem solving.
- Outstanding efficiency vs open-source models. GLM5-744B-A40B and Kimi-K2.5-1T-32B score only 16.6 and 20.1 on FLTEval; Leanstral surpasses both with just pass@1. Strongest OSS competitor Qwen3.5-397B-A17B achieves 25.4 at pass@4, while Leanstral hits 26.3 at pass@2.
- Dominant cost efficiency vs Claude family. Leanstral pass@2 ($36) scores 2.6 points higher than Claude Sonnet ($549), and pass@16 ($290) scores 8 points higher. Claude Opus 4.6 leads at 39.6 but costs $1,650 — 92x Leanstral's cost.
- Trained for maximum performance with lean-lsp-mcp (Lean Language Server Protocol wrapped as MCP), and can attach arbitrary MCPs using Mistral Vibe as scaffold.
- Real-world case: Leanstral solved a compilation issue in Lean 4.29.0-rc6 by self-diagnosing a definitional equality problem where a `def` type alias blocked `rw` tactic pattern matching, and proposed changing `def` to `abbrev`.
Evidence
- Formal verification was praised as a structural solution for AI coding. Verification suites accumulate as executable documentation expressing 'how code should behave' — consuming zero context tokens when code is correct, making them stronger than markdown specs.
- Counter-argument that formal verification doesn't solve AI code's fundamental problems. Functions matching specs can be proven, but security requirements nobody writes in specs — like 'don't hardcode database credentials,' 'don't leave CORS open,' 'add auth to admin routes' — are outside formal verification's scope.
- Cost-performance interpretation was questioned. Leanstral is 10x cheaper than Haiku but also lower performing — in accuracy-critical tasks, does 'cheaper but worse' matter? However, Opus not performing great on this benchmark was seen as hopeful, suggesting scaling Leanstral could potentially beat Opus.
- Skepticism about practicality. In real software shops, even property-based testing is hard to adopt, let alone formal proofs. Also noted that AI-generated proofs would be hard to read and inelegant, and the requirement for humans to verify that proofs are specifying the right thing doesn't disappear.
- Criticism of Mistral's strategic direction — focusing on non-mainstream academic areas while falling behind on frontier models. Counter-argued that companies like Mistral matter for model alignment diversity. Positive confirmation that weights are genuinely Apache 2.0 open-source.
How to Apply
- If repetitive proof writing is a bottleneck in Lean 4 math libraries or formal verification projects, connect Leanstral with lean-lsp-mcp in agent mode for comparable or better proof completion rates at 1/15th Claude Sonnet's cost at pass@2-4.
- When Lean version upgrades cause sudden compilation failures, have Leanstral write test code reproducing the failure environment first, then diagnose the cause — useful for tracking breaking changes without domain expertise.
- For mission-critical software needing accuracy guarantees with AI code generation, write core business logic specs in Lean 4 and build a pipeline where Leanstral proves implementations satisfy specs — reducing the human review bottleneck. Note: security requirements must be explicitly included in specs to be verifiable.
- Currently accessible via Mistral Vibe and free API endpoints, so you can experiment with Lean 4 at zero cost. The community also suggested testing whether ensemble passes across models (e.g., Leanstral → Qwen → Leanstral) outperform repeated same-model passes.
Terminology
Related Papers
Show HN: ctx – Search the coding agent history already on your machine
Claude Code, Cursor, Codex 등 코딩 에이전트가 이전 세션의 논의·결정·실패 시도를 잊지 않도록 SQLite로 인덱싱해 재사용할 수 있게 해주는 오픈소스 CLI 도구다.
Micro-Agent: Beat Frontier Models with Collaboration Inside Model API
vLLM 팀이 단일 모델 API 호출 뒤에서 여러 모델이 협업하는 'Micro-Agent' 개념을 공개했습니다. 별도의 에이전트 코드 없이 라우터 레이어에서 모델 조합을 실행해 GPT-4급 결과를 더 저렴하게 낼 수 있다는 아이디어입니다.
Ornith-1.0: self-improving open-source models for agentic coding
Gemma 4와 Qwen 3.5를 기반으로 파인튜닝한 코딩 특화 오픈소스 모델로, RL(강화학습)을 통해 스캐폴드(에이전트 실행 구조)까지 함께 최적화하는 방식을 주장하지만, 커뮤니티에서는 벤치마크 과최적화에 불과하다는 의심을 받고 있다.
Entity Binding Failures in Tool-Augmented Agents
AI 에이전트가 올바른 도구를 선택해도 잘못된 대상에 실행하는 'Entity Binding 실패' 문제를 정의하고, 이를 막는 실행 정책을 평가한 논문.
Herdr: Agent multiplexer that lives in your terminal
여러 AI 코딩 에이전트(Claude, Codex 등)를 하나의 터미널에서 동시에 실행·관리할 수 있는 Rust 기반 오픈소스 툴로, tmux처럼 세션이 유지되고 SSH로 원격 접속도 가능해 멀티 에이전트 워크플로우를 크게 단순화해준다.
Ornith-1.0: Self-scaffolding LLMs for agentic coding
모델이 문제 풀이 전략(scaffold)을 직접 생성하고 개선하는 자기강화 학습 프레임워크를 적용한 오픈소스 코딩 특화 LLM으로, 9B 소형 모델부터 397B 대형 모델까지 라인업을 갖추고 SWE-Bench 등 주요 벤치마크에서 Claude Opus 4.7을 능가하는 성능을 보여줬다.