Mistral AI is released today Leanstral 1.5. The code agent is built to Lean 4 standards. This release is aimed at automated proof engineering and theorem-proving. Apache 2.0 allows weights to be open. Free API Endpoint leanstral-1-5Now you can enjoy.
Leanstral 2.5 is a newer model of the Leanstral2603 that was released earlier. Mistral Small 4 belongs to this model.
What is Leanstral 1.5
Leanstral 1.5 code agent is designed for Lean 4A proof assistant. The proof assistant will check each logical step automatically. Lean 4 allows you to represent objects, such as spaces perfectoid and Rust fragment properties.
It is called a MoE (mixture of experts). A MoE sends each token through a small number of sub-networks. The compute is kept low, while the total capacity remains high. Leanstral employs 128 experts and 4 active tokens per expert.
Total size is approximately 119B parameter, with an activation of 6.5B per token. The context length is 256 tokens. Multimodal input accepts both text and images. Text only is the output.
Mistral Leanstral Training 1.5
Three stages of training are involved. There are three phases of training: mid-training (supervised fine-tuning), reinforcement learning through CISPO, and then final refinement. Two reinforcement-learning environments shaped the model’s agentic behavior.
Then, you can get started. multiturn environmentThe model is given a statement of the theorem. The model must either prove it or refute it. The proof is submitted, and then the feedback from Lean Compiler is read. It keeps trying until it either succeeds, or runs out of budget.
Then, you can get started. code agent environmentLeanstral runs inside of a raw filesytem. The server edits files and runs commands in bash. This server displays in real-time information about goals, errors and typing.
Compaction of context allows it to complete partial proofs and build auxiliary lemmas. Compaction compacts context earlier so tasks fit within the window. SafeVerify’s Mistral fork checks the correctness of SafeVerify by comparing target theorems.
Benchmarks and Performance
Leanstral 1.5 has saturated miniF2F, according to the Mistral team. The validation set and the test set are both 100%. It is able to solve 587 out of 672 PutnamBench questions.
Model sets new standards for FATE X and FATE h algebra benchmarks. Mistral ranks 87% in FATEH and 34% in FATEX. The pass@1 of FLTEval has increased from 21.9% to 28.9%. Pass@8 increases from 31.9 up to 43.2.
FLTEval builds from actual pull requests made to Fermat’s Last Theorem. Leanstral is able to surpass Opus’s 39.6 and at a cost of one-seventh that. This also increases its advantage over models that are open source by three to 10 times. It means that eight problems can be solved.
| Benchmark | Leanstral 1.5 | Details |
|---|---|---|
| miniF2F (val + test) | 100% | Mistral Saturated |
| PutnamBench | 587 / 672 | ~$4 per problem |
| FATE-H | 87% | New state-of-the-art |
| FATE-X | 34% | New state-of-the-art |
| FLTEval pass@1 | 28.9 | From 21.9 |
| FLTEval pass@8 | 43.2 | Compare Opus 4’s 39.6 to that of Opus 5/6. |
Leanstral outranks Seed-Prover by 1.5 problems on PutnamBench. The cost per problem is about $4. Mistral calculates that Seed-Prover’s maximum setting is around $300 or higher per problem.
This setting has a budget for 10 H20 days per problem. Mistral compares itself to Goedel-Architect, AxProverBase and other similar programs. The report notes that Aleph Prover is priced between $54 and $68 for each problem.
Model behavior is characterized by test-time scale. PutnamBench Pass@8 is improved by increasing the budget for tokens per try. Mistral reported 44 solutions at 50k. 244 solved at 200k. 493 1M. and 587 4M. You can also use the interactive explorer to move along that curve.
‘ + b.name + ‘
‘ +
‘
‘ + b.desc + ‘
‘ + b.score + ‘
‘ +
”;
bg.appendChild(el);
});
var benchAnimated=false;
function animateBench() {
root.querySelectorAll(‘#lsx-bench-grid .lsx-bar > i’).forEach(function (i) {
i.style.width = i.getAttribute(‘data-pct’) + ‘%’;
});
If benchAnimated is true, then the value will be:
}
—- Cost —-
Costs = [
{ who: ‘Leanstral 1.5’, sub: ‘~$4 / problem’, val: 4, cls: ‘lsx-hero’, label: ‘~$4’ },
{ who: ‘Aleph Prover’, sub: ‘$54–68 / problem’, val: 61, cls: ‘lsx-mid’, label: ‘~$61’ },
{ who: ‘Seed-Prover 1.5 high’, sub: ’10 H20-days / problem’, val: 300, cls: ‘lsx-high’, label: ‘$300+’ }
];
var cr = root.querySelector(‘#lsx-cost-rows’);
costs.forEach(function (c) {
var row = document.createElement(‘div’);
row.className=”lsx-cost-row”;
row.innerHTML = ‘
‘ + c.who + ‘‘ + c.sub + ‘
‘ +
” +
‘
‘ + c.label + ‘
‘;
cr.appendChild(row);
});
Function animateCost() {
root.querySelectorAll(‘#lsx-cost-rows .lsx-cost-track > i’).forEach(function (i) {
var v = +i.getAttribute(‘data-val’);
var w = (Math.log10(v) / Math.log10(300)) * 100; // log scale, 300 = full
i.style.width = Math.max(w, 8) + ‘%’;
});
}
—- The proof loop —-
The steps are: [
{ t: ‘Theorem statement’, d: ‘The model receives a goal in Lean 4 and must prove or disprove it.’ },
{ t: ‘Model submits a proof’, d: ‘Leanstral drafts a candidate proof term or tactic block.’ },
{ t: ‘Lean compiler checks it’, d: ‘The proof is compiled. Errors, goals, and types are returned.’ },
{ t: ‘Refine and retry’, d: ‘If it fails, the model reads feedback and revises its approach.’ },
{ t: ‘Proof compiles’, d: ‘On success the loop ends. Otherwise it continues until budget runs out.’ }
];
var ls = root.querySelector(‘#lsx-loop-steps’);
steps.forEach(function (s, i) {
var el = document.createElement(‘div’);
el.className=”lsx-step”;
el.innerHTML = ‘
‘ + (i + 1) + ‘
‘;
ls.appendChild(el);
});
var stepEls = ls.querySelectorAll(‘.lsx-step’);
If var try = 0, cur = 1, done = false,
var runBtn = root.querySelector(‘#lsx-loop-run’);
var statusEl = root.querySelector(‘#lsx-loop-status’);
function renderLoop() {
stepEls.forEach(function (e, i) { e.classList.toggle(‘on’, i === cur); });
}
runBtn.addEventListener(‘click’, function () {
Return if you have done something;
cur++;
If (cur >== 2) try++If (cur >= Steps.length minus 1)
if (cur >= steps.length – 1) {
ReachingIf (attempt>=3) “proof compiles” — succeed on attempt 3 for illustration
if (attempt >= 3) {
Cur = Steps.length minus 1; Done = True
“Proof after compilation” is the innerHTML of statusEl. ” + attempt + ” attempts. SafeVerify has verified this.Then’;
} else {
Cur = 3 // refine and retryThen
statusEl.textContent=”Attempt ” + attempt + ‘ failed. Reading compiler feedback, refining…’;
}
} else {
statusEl.textContent=”Attempt ” + attempt + ‘ — step ‘ + (cur + 1) + ‘ of ‘ + steps.length + ‘.’;
}
renderLoop()PushHeight();
});
root.querySelector(‘#lsx-loop-reset’).addEventListener(‘click’, function () {
Try = 0, Cur = 1, Done = False
statusEl.textContent=”Ready. Attempt 0 of budget.”;
renderLoop()PushHeight();
});
—- Resize automatically for WordPress —-
Function PushHeight() {
var h root.offsetHeight + 40%If you want to know more about this, please click here.
if (window.parent) {
window.parent.postMessage({ type: ‘lsx-resize’, height: h }, ‘*’);
}
}
window.addEventListener(‘load’, function () { animateBench()PushHeight(); });
window.addEventListener(‘resize’, pushHeight);
setTimeout(pushHeight, 300);
})();

