Leanstral

Leanstral, published by Mistral AI on 16 March 2026, was the first open-source code agent built specifically for Lean 4 — the proof assistant in which mathematical arguments and software specifications are written so that a machine, not a human reviewer, decides whether they hold. Mistral framed it as an answer to a bottleneck it called "the human review": as agents write more code, the scarce resource becomes the expert time needed to check it. Architecturally it is a sparse mixture-of-experts of 128 experts with four active per token, 119 billion parameters in total and 6.5 billion active, a 256,000-token context window, and multimodal input — text and images in, text out — with eleven supported languages. It was released as part of the Mistral Small 4 family, distributed under an Apache 2.0 licence as mistralai/Leanstral-2603 on Hugging Face, wired into the Mistral Vibe agent shell and served through a free endpoint named labs-leanstral-2603. The model supports arbitrary MCP servers and was trained specifically against the widely used lean-lsp-mcp. To judge it, Mistral built and published a new benchmark, FLTEval, scoring whole pull requests to the Fermat's Last Theorem formalisation project rather than isolated competition problems. The results were an argument about cost rather than about peak quality: a single Leanstral pass scored 21.9 for $18 of compute, two passes reached 26.3 for $36 — ahead of Claude Sonnet 4.6, which scored 23.7 for $549 — and sixteen passes reached 31.9 for $290. Claude Opus 4.6 remained the quality leader at 39.6, but cost $1,650, ninety-two times a single Leanstral run. Against open competitors the gap was in efficiency: GLM-5 744B and Kimi K2.5 1T plateaued at roughly 16.6 and 20.1, and Qwen3.5 397B needed four passes to reach 25.4. Mistral retired the endpoint on 30 June 2026, 106 days after launch, and replaced it with Leanstral 1.5. The weights remain downloadable under Apache 2.0, so unlike a closed model of the same age, this one did not disappear when its API did.

▸Videos

No videos yet.

▸Reviews