Mistral AI has introduced a large language model called Devstral, aimed at application development (vibe coding) and optimized for formal code verification. It is expected that Devstral can be used to create AI assistants that not only generate code but also ensure its correctness.
Devstral is the first open model to support the Lean 4 programming language and its associated tools for verifying mathematical proofs. Lean 4 provides capabilities for mathematically proving the correctness of code and its compliance with specifications, enabling confirmation in vibe coding that the code generated by the AI model does exactly what is intended.
The model encompasses 119 billion parameters (6.5 billion activatable parameters per token), considers a context of 256 thousand tokens, and is released under the Apache 2.0 license. The downloadable archive with Leanstral is 121 GB and is suitable for use on local systems. Libraries such as vllm, transformers, and SGLang can be applied for local deployment.
Among other uses, the model can be utilized for vibe development in the open agent mistral-vibe and can integrate with the Aeneas toolkit for code verification in Rust. The input parameters include text and images, with text as the only output. Image content analysis is supported.
To assess the capabilities of AI models regarding the quality of formal code verification and the writing of mathematical proofs, a new set of tests called FLTEval has been developed. In the conducted tests, the Leanstral model significantly outperformed existing open models Qwen3.5 397B-A17B, Kimi-K2.5 1T-A32B, and GLM5 744B-A40B, showing comparable results with Anthropic's Claude Haiku 4.5 and Claude Sonnet 4.6 models, but fell behind Claude Opus 4.6. Specifically, the Opus model scored 39.6 points, while Leanstral scored 21.9 on a single pass and 31.9 on 16 passes. The costs for using Opus were $1650, while Leanstral cost $18 on a single pass and $290 on 16 passes. The Haiku model scored 23 points at a cost of $184, while the Sonnet model scored 23.7 at a cost of $549.
Source: opennet.ru
