Leanstral 1.5 review 2026. Honest scoring on Lean 4 theorem proving, autoformalization, the 119B/6.5B MoE, 256k context, free Labs pricing, and who it is for.
Leanstral 1.5 is a strong, focused Lean 4 proof model: efficient MoE architecture, a big 256k context for long proofs, and free access in Mistral Labs. It does one hard thing — propose machine-checkable proofs and formalize specs — and does it well. The main caveats are honesty caveats: Mistral has not published fresh benchmarks for 1.5, and weight availability is unconfirmed, so score it on the category it serves and verify the specifics before betting a project on it.
Leanstral 1.5 arrived on June 30, 2026 as Mistral's update to the original Leanstral (March 2026), and it is a narrow tool with a clear job. It is a formal proof engineering model for Lean 4 — you use it for automated theorem proving and autoformalization, not for writing, chat, or general reasoning. Framing it as a "language model for efficient text generation" undersells and mislabels it; the efficiency is real, but the target is formal proofs.
The specs are concrete. It is a Mixture-of-Experts model with 119B total parameters and 6.5B active per token, paired with a 256k-token context window sized for long proof files, and it is listed at $0 on Mistral's Labs tier via the console playground. What is missing is a fresh benchmark story: Mistral has not published new head-to-head numbers for 1.5, and the reference point remains the original Leanstral, which the company said beat Claude Sonnet by 8 points at pass@16 while costing roughly 15x less.
This review scores Leanstral 1.5 for what it is — a specialized proof model — not as a general assistant or a content tool, because it is neither. If you work in formal methods or verified software, read on for where it is strong and where the uncertainty sits. If you are a creator who found this while wondering whether it writes posts, the short answer is no, and the honest positioning is at the end.
Leanstral 1.5 is a formal proof engineering model from Mistral, the French AI lab. It targets two tasks in Lean 4, a proof-assistant language: automated theorem proving, where you supply a proof goal and the model proposes steps that the Lean checker mechanically verifies, and autoformalization, where it converts human-readable math or a software specification into formal Lean 4 definitions and theorems. The value is that the output is verified rather than merely plausible — a checker either accepts the proof or it does not. Architecturally it is a Mixture-of-Experts model (119B total parameters, 6.5B active per token) with a 256k-token context window. It is served through a chat-style API supporting chat completions, function calling, and agent workflows, and is offered free on Mistral's Labs tier as labs-leanstral-1-5. It is the successor to the March 2026 Leanstral, which shipped open weights under Apache 2.0 on Hugging Face; for 1.5, Mistral has confirmed only the free Labs access so far.
The clear fit is a researcher, mathematician, or engineer doing formal verification — proving theorems, formalizing specifications, or building verified software in domains where a single bug is expensive, like cryptography, aerospace, and financial modeling. Teams already working in Lean 4 who want an efficient, free proving assistant get the most out of it. It is a poor fit for anyone outside formal methods: it is not a general writing or coding assistant, not a chatbot, and not a content tool, and using it as one wastes what makes it good.
| Dimension | Score | Why |
|---|---|---|
| Theorem-proving capability | 4.2 / 5 | Proposes machine-checkable Lean 4 proofs; the prior release was reported to beat Claude Sonnet by 8 points at pass@16, though 1.5 lacks published benchmarks. |
| Autoformalization | 4.0 / 5 | Converts human-readable math and specs into Lean 4 definitions and theorems — a genuinely hard capability few models target. |
| Model efficiency (MoE) | 4.5 / 5 | 119B total parameters with only 6.5B active per token keeps inference cost low relative to the parameter pool. |
| Context window | 4.5 / 5 | 256k tokens is large enough to hold long proof files and multi-file developments. |
| Pricing / value | 4.5 / 5 | Listed at $0 on Mistral Labs — free to try, which is rare for a specialized model of this size. |
| Openness / deployment | 3.5 / 5 | The original Leanstral was Apache 2.0 open weights; 1.5 confirms only free Labs access, so self-hosting status is unconfirmed. |
| Benchmark transparency | 3.5 / 5 | No new head-to-head benchmarks published for 1.5 at launch; the numbers on record belong to the March release. |
| Ease of use for non-experts | 3.0 / 5 | Requires working knowledge of Lean 4 and formal methods; there is no gentle on-ramp for a general user. |
Leanstral 1.5 is listed at $0 on Mistral's Labs tier, which makes the value calculation simple for experimentation: it costs nothing to try in the console playground. For a researcher or team evaluating whether a specialized proof model can accelerate their Lean 4 work, free access removes the usual barrier to a serious trial, and the original Leanstral's reported economics — competitive proof performance at roughly 15x less cost than a frontier model at pass@16 — suggest the line is designed around cost-efficiency.
The caveat is that "free in Labs" is not the same as a committed, priced production tier. Mistral has not published a paid Leanstral 1.5 tier at launch, and it has not confirmed whether 1.5 ships open weights the way the original did under Apache 2.0. For a hobby project or a research experiment, none of that matters. For a team that wants to build a verification pipeline on top of it, the open question is what production access and licensing look like beyond Labs — worth confirming with Mistral before committing.
Measured against pointing a general frontier model at the same Lean 4 problems, a free specialized model is an easy win on cost, provided it holds up on your proofs. As always with a narrow tool, the real value is throughput on your specific workload, so validate it on your own goals rather than on the prior release's benchmark numbers.
| Use case | Fit | Why |
|---|---|---|
| Automated theorem proving in Lean 4 | Strong | This is Leanstral's core purpose — proposing proof steps a checker verifies. |
| Autoformalizing math or specifications | Strong | It converts human-readable statements into formal Lean 4 definitions and theorems. |
| Verified software in safety-critical domains | Strong | Cryptography, aerospace, and financial modeling benefit from machine-checked correctness. |
| Free experimentation with a proof model | Strong | $0 Labs access lets you trial it without procurement. |
| Building a production verification pipeline | OK | Capable, but Labs-only access and unconfirmed 1.5 weights mean you should confirm production terms first. |
| General coding assistance | Weak | It is specialized for formal proofs, not everyday code generation. |
| Writing, chat, or content generation | Weak | It does not write prose; it produces proofs. Use a general model or a content engine instead. |
Kompozy is not a competitor to Leanstral 1.5 and it would be dishonest to score it as one — they do not touch the same job. Leanstral proves things; Kompozy makes and publishes content. The only reason to mention it in a Leanstral review is that the two sit at opposite ends of one workflow that technical teams keep running into: you prove the result, and then someone still has to explain it to the world.
For a DevRel lead or a technical educator evaluating both, the honest split is clean. Leanstral is the tool your engineers use to produce a verified proof or a formalized spec. Kompozy is the tool your marketing or developer-advocacy function uses to turn the plain-language write-up of that work into a carousel, a blog, an X thread, a newsletter, and a short persona video — in your brand voice via the Persona Brief, scheduled across nine platforms. If your gap is proving, Kompozy is irrelevant. If your gap is that great formal-methods work keeps going unseen, that is exactly what Kompozy addresses, and it pairs with Leanstral rather than replacing it.
If you do Lean 4 proof work, yes — it is a capable, efficient, specialized model and it is free to try in Mistral Labs. The caveats are that Mistral has not published new benchmarks for 1.5 and has not confirmed open weights or a production tier, so validate it on your own proofs and confirm access terms before building a pipeline on it.
Two things: automated theorem proving (proposing Lean 4 proof steps that the checker verifies) and autoformalization (turning human-readable math or software specifications into formal Lean 4 definitions and theorems). It is a proof model, not a general assistant or a content tool.
Mistral lists it at $0 on its Labs tier, free to try in the console playground as labs-leanstral-1-5. There is no published paid tier at launch, and Labs availability can change, so check Mistral's console for current terms.
It is the June 30, 2026 update to the March 2026 Leanstral, described as an optimised Lean 4 proof-engineering model. Mistral has not published new head-to-head benchmarks for 1.5; the numbers on record — beating Claude Sonnet by 8 points at pass@16 at roughly 15x less cost — belong to the original release.
The original Leanstral shipped open weights under Apache 2.0 on Hugging Face. For 1.5, Mistral has confirmed only the free Labs access at launch, so open-weight and self-hosting status for this version is unconfirmed. Verify before relying on it.
No. Despite being served through a chat-style API, it is tuned for Lean 4 proofs and formalization, not writing, chat, or content. To turn a verified result into posts, blogs, or video and publish them, use a content engine like Kompozy downstream.
For proof work: the original Leanstral (open weights), general frontier models pointed at Lean 4, dedicated open proof models like the DeepSeek-Prover line, or a manual Lean 4 plus Mathlib workflow. For turning your verified work into content, Kompozy — a different category, used downstream.