Mistral ships Leanstral 1.5 for Lean 4 proof work, free in Labs

Lean 4 has been quietly turning into the venue where frontier labs argue about reasoning, and Mistral just dropped a refresh that sits squarely in that lane. Mistral's model card describes Leanstral 1.5 as "An updated Lean 4 formal proof engineering model optimised for automated theorem proving and autoformalization," with 119 billion total parameters, 6.5 billion of them active per token, and a 256k-token context window. The Labs listing puts the price at $0 and points users at the playground on console.mistral.ai.

The lineage is worth knowing. When Mistral shipped the original Leanstral in March, the pitch was that this exact sparse mixture-of-experts shape, with each token routed to just 4 of 128 expert modules, could outperform much larger open models on formal proofs while staying cheap to run. The company's reported numbers had it beating Claude Sonnet by 8 points at pass@16 while costing 15x less, with Anthropic's Opus still ahead on quality but at 92x the cost. 1.5 is presented as an update to that same model, not a new architecture.

Why a refresh of a niche math model matters: formal proof tooling is where "the answer was wrong but it sounded right" stops working, because Lean checks the proof. A Lean-fluent model that runs on 6.5B active parameters is potentially useful well beyond pure mathematics, because the same shape underwrites verified-software work, where teams want code that ships with a machine-checkable correctness argument attached. Mistral is giving the latest version away for free in the playground, which lowers the cost of taking it for a spin.

The honest caveat is that the new model card does not publish fresh benchmark numbers against the March 2603 release or against the bigger systems chasing the same problem, so "how much better is 1.5" is unanswered today. The other open piece is weights and licensing: the prior Leanstral was published on Hugging Face under Apache 2.0, but the 1.5 page only confirms the free Labs tier. For people working in Lean, autoformalization, or verified-code agents, this is the version to put in front of your test harness next, and worth watching for an updated leaderboard once the post-release write-ups land.