0.01%. That is roughly the fraction of the open web that consists of formally verified mathematical proofs. While the internet is drowning in “math”—blog posts, textbooks, and StackExchange threads—almost all of it is natural language. It is descriptive, imprecise, and frequently wrong. Mistral is betting that the only way to move past the “vibes” era of LLM reasoning is to stop training on what humans say about math and start training on what can actually be proven.
Mistral is admitting something that most labs try to hide: the web is a wasteland for high-quality logic. If you want a model that doesn’t hallucinate a fake lemma halfway through a proof, you can’t just feed it more Common Crawl. You have to build a factory. By focusing on synthetic proof generation, Mistral is essentially creating a closed-loop system where the model proposes a proof and the Lean compiler acts as the ultimate judge. If the compiler says no, the data is discarded. It is a brutal filtration process that ensures the training set is logically perfect. This is the core philosophy behind Leanstral 1.5, and it represents a shift from predicting the next token to predicting the next valid logical step.
This approach treats mathematical reasoning like a compiler problem rather than a translation problem. Most LLMs treat math as a sequence of tokens that look like a solution; Leanstral treats it as a program that must execute without errors. It is like the difference between a chef describing how to bake a cake and a chemist providing the exact molecular formula for the ingredients. One is an approximation; the other is a specification. Who is actually going to use this? (Probably the same people who enjoy spending their weekends in a debugger). But for those people, the ability to automate the grunt work of formal verification is a massive win. It removes the cognitive load of the boilerplate and lets the human focus on the high-level strategy of the proof.
The strategic play here is a clear pivot away from the “everything app” mentality of OpenAI. While the industry giants are fighting over who can build the best general-purpose assistant for writing emails, Mistral is carving out a space as the lab for people who actually care about the underlying logic. They are building tools for the engineers and mathematicians who find the current “reasoning” models too fluffy. However, there is a catch: the friction here isn’t the model, but the tool. Lean 4 has a learning curve that looks like a vertical cliff (and it is a brutal one), which limits the immediate utility of the model to a very small circle of specialists. Or maybe not—perhaps this is exactly how you build a moat.
Still, the synthetic data approach is the real story. If they can successfully generate “abundance” from a small seed of formal logic, they’ve found a way to scale intelligence without needing more human-written text. We have reached the limit of what the public web can provide. The future of reasoning is synthetic, and it must be verified by an external source of truth—whether that is a compiler, a physics engine, or a game of chess. I suspect we will see a competing “formal-first” model from another major lab by Q4. The race is no longer about who has the biggest dataset, but who can build the best filter to create a clean one.
A niche tool for a niche crowd, but a vital one.
