Not established from the available sources.
MISTRAL AI · MODEL RELEASE TRACKER
Leanstral-120B-A6B
Leanstral-120B-A6B is an open-source code agent released by Mistral for the Lean 4 proof assistant. It is optimized for proof engineering and formal repositories, uses a highly sparse architecture with 6B active parameters, and is distributed under an Apache 2.0 license via Mistral vibe and a free API endpoint.CURRENT SNAPSHOT2/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
Not established from the available sources.
Benchmark / evaluationBenchmarked using a new FLTEval suite focused on completing all formal proofs and correctly defining new mathematical concepts in each PR to the FLT project; compared against several leading coding agents (Claude Opus 4.6, Sonnet 4.6, Haiku 4.5) and open-source models (Qwen3.5 397B-A17B, Kimi-K2.5 1T-A32B, GLM5 744B-A40B).
License and accessWeights released under an Apache 2.0 license; available in an agent mode within Mistral vibe and accessible via a free API endpoint.
VERIFIABLE FACTS
Every value stays attached to a source and date.
RELEASE · ReleaseDEVELOPER CLAIM
Released by Mistral on 2026-03-16 as the first open-source code agent designed for the Lean 4 proof assistant.
CAPABILITIES · Active parametersDEVELOPER CLAIM
Uses 6B active parameters (described as highly efficient with 6B active parameters).
CAPABILITIES · Architecture and inferenceDEVELOPER CLAIM
Built with a highly sparse architecture, optimised for proof engineering tasks and leveraging parallel inference with Lean as a verifier.
CAPABILITIES · Training focus and MCP supportDEVELOPER CLAIM
Trained to operate in realistic formal repositories and for proof engineering; supports arbitrary MCPs through Mistral vibe and was specifically trained for strong performance with the lean-lsp-mcp.
AVAILABILITY · License and accessDEVELOPER CLAIM
Weights released under an Apache 2.0 license; available in an agent mode within Mistral vibe and accessible via a free API endpoint.
BENCHMARKS · Benchmark / evaluationDEVELOPER CLAIM
Benchmarked using a new FLTEval suite focused on completing all formal proofs and correctly defining new mathematical concepts in each PR to the FLT project; compared against several leading coding agents (Claude Opus 4.6, Sonnet 4.6, Haiku 4.5) and open-source models (Qwen3.5 397B-A17B, Kimi-K2.5 1T-A32B, GLM5 744B-A40B).
WHAT CHANGED