Mistral Small 4: An AI Model with 119 Billion Parameters

Mistral Small 4: A Major Technological Advancement

Mistral recently announced the launch of its latest artificial intelligence model, the Mistral Small 4. This model, licensed under Apache 2, boasts 119 billion parameters thanks to a Mixture-of-Experts architecture, of which 6 billion are actively used. This advancement marks a significant milestone in the field of AI.

The Mistral Small 4 stands out for its ability to integrate the features of three flagship models from Mistral: Magistral, focused on reasoning, Pixtral, specialized in multimodal capabilities, and Devstral, designed for agentic coding. This integration makes it a versatile model capable of addressing various needs.

Features and Capabilities

The model offers two levels of reasoning effort: reasoning_effort="none" and reasoning_effort="high", the latter providing verbosity comparable to previous models of Magistral. Mistral Small 4 is available on the Hugging Face platform, where it occupies 242 GB.

During tests conducted via the Mistral API, the model demonstrated its ability to generate complex content, although the current documentation does not yet allow for the definition of reasoning effort, a feature expected in future updates.

Leanstral: A Model for the Lean 4 Language

In parallel, Mistral has also launched Leanstral, an open-weight model specifically designed for the formally verifiable coding language Lean 4. Although the author has not explored Lean in depth, Mistral's initiative to target a specific language underscores their commitment to diversifying AI applications.

These developments reflect Mistral's dedication to pushing the boundaries of artificial intelligence by providing powerful tools for reasoning, multimodality, and coding.