Summary: Mistral open-sourced Leanstral-120B-A6B, its first code agent built for Lean 4 proof engineering, under Apache 2.0. The sparse model is positioned for repository-level proofs and can be self-hosted, which matters to formal-verification teams that need code to remain on their machines. The Hacker News discussion welcomed the direction while reserving judgment against Qwen until independent benchmarks appeared.
Leanstral is here: Mistral has handed the open-source community its first code agent designed from the ground up for Lean 4. The model carries the name Leanstral-120B-A6B — a highly sparse design holding 120B parameters in total, of which about 6B stay active — and its weights are published in full under Apache 2.0 (the Hugging Face model card, mistralai/Leanstral-2603, prints the spec as 119B-A6B). This is not being pitched as a demo toy; the assignment is realistic proof engineering — walking into complete repositories to finish proofs and define concepts, rather than cracking isolated exercise problems. For teams practicing formal verification, a same-class tool can finally be installed and run on their own machines, and that is a first.
A Lean 4 Specialist Arrives in Open Source
A conventional code agent chases code that merely looks correct; Lean 4 recognizes only two outcomes, machine-verified or not true. Mistral pointed its agent straight at that high-bar territory: training ran against lean-lsp-mcp, and because any MCP server is supported, the agent can genuinely read and mutate the state of a Lean project rather than having to infer everything from a frozen snapshot of the file. The framing of the work is franker than most agent marketing as well — finishing proofs and defining new concepts inside full repositories is a different, harder job than one-shotting isolated lemmas, and it is precisely the job real formal-verification teams carry. The Hacker News thread showed the familiar reception: broad approval for the direction, with judgment on how it measures against the Qwen family deliberately held back until independent benchmarks appear on someone else’s hardware.
What Makes Lean 4 Worth Targeting
Lean 4 works as a proof assistant. It can capture frontier mathematical objects, perfectoid spaces among them, and it can just as well state software specifications — verified properties of Rust code being one case. Its value lies in the exchange it enables: swapping the bottleneck of “humans reviewing machine-generated code” for “machines proving it”. Once a proof goes through, correctness stops being a probability and becomes a mathematical fact. That trait explains why formal methods have long lived in compilers, kernels, and cryptography — settings where a bug costs enough to make the rigor worthwhile. With the volume of AI-generated code exploding, the review bottleneck scales in step with output while trust does not, and formal verification is moving from a mathematician’s luxury to a requirement for high-stakes software. That is exactly the bet Leanstral places: if machines are going to write the code anyway, machine-checking is the only review pipeline that can keep pace.
What 120B Total and 6B Active Do to the Economics
Sparsity in this product is the operating logic itself, not a discount slogan. Proof search is, at bottom, enormous parallel trial, and repeated-sampling tactics (pass@k) only stand a chance when a single inference is cheap: $18 covers one pass, and sixteen passes come to $290. Put Claude Opus beside it at $1,650 for a single run — 92 times the price of one Leanstral pass, and that is before anyone has tried a second angle on the proof. The ledger tilts further toward self-hosting: with the model living on hardware you own, another sampling round costs electricity and GPU time, not a metered token bill. The arithmetic Mistral is playing is straightforward — a cheap, parallel-iterable specialized model earns the footing to take on general-purpose flagships inside a vertical where iteration count outweighs raw single-shot intelligence.
Making Sense of the FLTEval Numbers
FLTEval is scored against real pull requests from the FLT project, and the Anthropic family draws the reference line: Claude Haiku put in $184 to reach 23.0, Sonnet $549 for 23.7, and Opus $1,650 for 39.6. Leanstral took 21.9 on a single $18 pass; pass@2 ($36) carried it to 26.3, past Sonnet at roughly one-fifteenth of the spending; pass@16 ($290) brought it to 31.9, eight points over Sonnet and still below the price of one Haiku run. Around it, the open-source field: Qwen3.5-397B-A17B required four passes to arrive at 25.4, Kimi-K2.5-1T-A32B finished near 20.1, and GLM5-744B-A40B near 16.6 — on this benchmark, specialization plainly delivered. Two caveats deserve a place in the reading: FLTEval sits close to Leanstral’s home ground, being built from the FLT project the model was developed around, and Opus’s 39.6 remains the outright ceiling. The fair summary is that Opus keeps the higher score while Leanstral takes cost per point by an order of magnitude — and at proof-search scale, that is frequently the contest that counts.
Where to Find It and How to Run It
Three ways in. The first is Mistral Vibe with no setup required: enter the /leanstall command, press Shift+Tab to pick Leanstral, or launch with vibe —agent lean. The second is the API: the labs-leanstral-2603 endpoint costs nothing or close to nothing for a limited window, which keeps experimenting against your own proof backlog inexpensive. The third is self-hosting: fetch the Apache 2.0 weights and run them on your own metal — Mistral’s own phrasing, and no minimum GPU is spelled out, though a 120B-parameter sparse model is clearly not a laptop workload. For regulated environments, proprietary-math teams, or anyone who will not let code leave the premises, the third option is the one that matters — and it is exactly why open weights remain a scarce, differentiated asset in 2026.