Mistral AI Revolutionizes AI with Small 4, Leanstral, and Forge

Mistral Small 4: A Hybrid and Dynamic Model

Mistral AI recently unveiled its latest model, Mistral Small 4, which stands out due to its hybrid architecture optimized for various applications such as general chat, coding, agentic tasks, and complex reasoning. This model is primarily aimed at developers, researchers, and businesses, integrating capabilities that were previously spread across multiple distinct models: text generation, reasoning, and image analysis.

Based on a Mixture of Experts (MoE) architecture, Small 4 features 128 experts, of which only 4 are activated per token, allowing for efficient scaling and specialization. It contains 119 billion parameters and offers a context window of 256,000 tokens. Mistral AI announces a 40% reduction in latency and a throughput three times higher than that of Small 3.

The fully open-source model allows users to dynamically adjust its behavior through the reasoning_effort parameter. Set to "none," it provides quick responses similar to Small 3, while on "high," it offers deeper reasoning. Small 4 is available under the Apache 2.0 license via the Mistral API, AI Studio, and Hugging Face.

Leanstral: An Innovative Coding Agent

In parallel, Mistral AI launched Leanstral, the first open-source coding agent designed for Lean 4. This formal proof assistant is capable of verifying complex mathematical and software specifications. Leanstral aims to enable developers to delegate not only code generation but also the demonstration of its compliance with expected specifications.

Leanstral is integrated into Mistral Vibe and is immediately accessible without additional configuration. Users can test it simply by using the /leanstall command.

Forge: A Tailored Platform for AI

Mistral AI also introduced Forge, a platform designed to enable businesses to create cutting-edge AI models from their own data and knowledge. Forge covers three phases of the development cycle: pre-training on large volumes of proprietary data, post-training to refine behavior on specific tasks, and reinforcement learning to align models with internal policies.

This platform allows organizations to train models that understand their internal context, thereby optimizing performance according to their specific constraints. Forge supports both dense and MoE architectures, as well as multimodal inputs. Already utilized by partners such as the European Space Agency, Ericsson, ASML, and Reply, access to Forge is available upon request via a contact form.