Key takeaways
- Leanstral 1.5 is Mistral AI’s new model for writing and checking Lean 4 math code.
- It is open under the Apache 2.0 licence, so developers can use it with fewer limits.
- On PutnamBench, the model solved 587 of 672 problems, or about 87.4%.
- The release matters because math-proof AI can help researchers, students, and coding teams verify hard logic.
Leanstral 1.5 is a new AI model from Mistral AI built for Lean 4. Lean 4 is a proof assistant, which means software that checks if a math proof is correct. Mistral says the model solved 587 of 672 PutnamBench problems, so this launch puts fresh focus on open math AI tools.
That headline sounds niche, but the idea is simple. Most chatbots write words. This model writes formal proofs, which are step-by-step math arguments a computer can test. Because the computer checks each step, the result can be more reliable than a normal text answer.
What is Leanstral 1.5 and why does it matter?
Mistral AI released Leanstral 1.5 as a code agent model for Lean 4. A code agent model is an AI system tuned to solve tasks with code-like steps. In this case, those steps are math proofs written in a strict language.
Lean 4 is popular in formal mathematics and software verification. Verification means checking that something is correct by rule, not by guess. So if an AI can work well in Lean 4, it may help people test the truth of hard claims.
This matters beyond math contests. Formal proof tools can help with chips, software, and security, because small mistakes can cause big failures. In fact, companies and labs use proof systems to check code that must not break.
Mistral also chose the Apache 2.0 licence. That is a widely used open-source licence, which means people can build with it and modify it under clear rules. For startups and researchers, that can matter almost as much as the raw benchmark score.
How strong is Leanstral 1.5 on PutnamBench?
The big number in the release is 587 solved problems out of 672 on PutnamBench. PutnamBench is a benchmark, which means a standard test used to compare models. It is based on hard university-level math problems in formal proof form.
That works out to about 87.4%. Here is the simple math: 587 divided by 672. A score near nine out of ten is striking, because these are not easy school exercises.
Leanstral 1.5 on PutnamBenchSolvedUnsolved58785Total problems: 672Solve rate: 87.4%
The unsolved count is 85. That gap is useful, because it shows where work remains. Even strong models still miss some hard cases, especially when proofs need deep insight or a long chain of exact steps.
| Metric | Value |
|---|---|
| Problems solved | 587 |
| Total problems | 672 |
| Unsolved | 85 |
| Solve rate | 87.4% |
| Licence | Apache 2.0 |
Readers should treat benchmarks with care, though. A benchmark score shows performance on one test set. It does not prove the model is best at every real task. Still, strong benchmark results often signal real progress.
Why are proof models different from normal chatbots?
Most AI chatbots predict likely text. They can sound smart even when they are wrong. Proof models face a tougher standard, because the proof checker accepts a step or rejects it.
That changes the game. The model cannot just write something that feels right. It has to produce code that matches formal rules exactly, a bit like solving a maze where every turn must fit.
So Leanstral 1.5 could be useful in places where trust matters. For example, a student could test a theorem idea. A researcher could draft proof steps faster. An engineer could check logic in safety-critical software.
There is still a limit. A proof assistant checks only what you ask it to check. If the original problem is framed badly, the software may still confirm the wrong thing. So humans still need judgment.
What does the open licence mean for developers?
The Apache 2.0 licence is a major part of this story. It is permissive, which means people usually get broad rights to use the software. That can make adoption faster in companies, labs, and open communities.
Many AI teams now care about two things at once: quality and access. A closed model may score well but stay hard to inspect. An open model gives developers more room to test, fine-tune, and plug it into their own tools.
That open angle fits a wider AI race. We have already seen heavy investment in enterprise AI, like Microsoft’s growing AI division. We have also seen how global names are shaping AI policy, as in the AI for Good Global Commission story.
For Indian developers, open models can be practical. Teams can test them locally, control costs, and build niche tools. That matters for startups that cannot spend huge sums on every API call. An API call is a paid request sent to another company’s AI service.
Where could Leanstral 1.5 be used next?
The near-term use is clear: formal math and theorem proving. But there are adjacent areas too. These include code verification, research assistants, and teaching tools that show each proof step clearly.
There is also a bigger industry angle. AI is moving from flashy chat to systems that can verify work. That shift may be slow, but it is important, because businesses often value correct answers more than pretty ones.
Mistral’s release may also push rivals to improve open proof models. Competition tends to speed up progress. Meanwhile, the real test will be whether people can use Leanstral 1.5 in day-to-day workflows, not just benchmark demos.
One simple way to think about it is this: if a regular chatbot is like a student talking through an answer, a proof model is like a student who must show every line of working. That is harder, but also much more useful when the stakes are high.
Leanstral 1.5 matters because it does more than chat about math. It writes Lean 4 proofs that software can check step by step, which makes its answers easier to trust on tightly defined tasks.
What should readers watch now?
First, watch for independent testing. Mistral shared the headline score, but outside researchers will want to confirm how the model performs in other settings. You can track official details on Mistral AI and the technical ecosystem around Lean.
Second, watch developer adoption. If proof researchers and coding teams start using Leanstral 1.5 widely, that will matter more than a launch-day chart. Tools that save time tend to spread fast.
Third, watch how this fits the open-model trend. We have seen other tech stories turn on scale and supply chains, from Tata Electronics’ iPhone export rise to platform battles over data and control. AI will likely follow the same pattern: better tools, tighter competition, and bigger fights over who gets to build on top.
Right now, the simple takeaway is this. Leanstral 1.5 looks like a serious new open model for formal math work. It will not replace mathematicians, but it could become a strong helper for people who need proofs that a machine can verify.
FAQs
What is Leanstral 1.5?
Leanstral 1.5 is an AI model from Mistral AI built to write and solve formal proofs in Lean 4.
How good is Leanstral 1.5 at math proofs?
Mistral says it solved 587 of 672 PutnamBench problems. That is about 87.4%.
Why does the Apache 2.0 licence matter?
It makes the model easier to use and adapt. So developers and companies may adopt it faster.
