French AI startup Mistral AI has introduced two new models, taking another step toward building AI that feels more useful in day-to-day work rather than just powerful on paper.
The company’s latest releases—Leanstral and Mistral Small 4—focus on solving practical problems, especially around trust, speed, and cost.
Leanstral is the more specialised of the two. It’s built for developers who need absolute correctness in their code. Designed for the Lean 4 environment, the model doesn’t just generate code—it checks and proves that the output actually meets the given requirements.
That’s important because, in most cases today, AI-generated code still needs to be reviewed and fixed by humans. It slows things down, especially in complex or high-risk projects. Leanstral changes that flow a bit. Instead of writing first and verifying later, developers can describe what they want, and the system handles both the coding and the validation together.
Technically, it runs on a sparse architecture with six billion active parameters, which helps keep it efficient. Lean itself acts like an internal checker, making sure the output is correct as it’s created. Mistral has released the model under the Apache 2.0 license, along with API access and integration into its Vibe coding platform. There’s also a new benchmark, FLTEval, meant to test how well such systems perform in real engineering scenarios.
In early tests, Leanstral performed quite well. It managed to beat several larger open-source models in single runs and stayed competitive even against systems that needed multiple attempts. It also showed a noticeable cost advantage compared to some proprietary options.
Alongside this, Mistral AI launched Mistral Small 4, which is more of an all-rounder.
Instead of using separate tools for different tasks, this model combines reasoning, coding, and image understanding into one system. It’s built using a mixture-of-experts design with 128 experts, with four active at a time, and supports a large context window of up to 256,000 tokens.
One practical feature here is control. Users can adjust how much reasoning the model uses—keeping it quick for simple tasks or letting it think a bit deeper when needed.
In internal benchmarks, Mistral Small 4 performed at a similar level or better than GPT-OSS 120B across several tests, while producing shorter outputs. That means faster responses and lower costs. On coding benchmarks like LiveCodeBench, it also outperformed competing models while generating around 20% less output.
Overall, these two models show where Mistral is heading—toward AI that is not just capable, but also dependable and easier to use in real-world situations.
Also Read: Mistral AI Acquires Koyeb in First Deal to Advance Its Cloud Ambitions
