Mistral AI Releases Unified Mistral Small 4 Model
- •Mistral Small 4 unifies reasoning, multimodal, and agentic coding capabilities into a 119B parameter model.
- •Model uses Mixture-of-Experts architecture with 6B active parameters under an Apache 2.0 license.
- •Release includes Leanstral, an open-weight model specialized for the Lean 4 formal verification language.
Mistral AI has significantly expanded its open-weights catalog with the release of Mistral Small 4. Despite its "Small" moniker, the model boasts 119 billion parameters, though it utilizes a Mixture-of-Experts (MoE) architecture to keep computational costs low by only activating 6 billion parameters during inference. This release marks a strategic shift for the French startup, as it consolidates capabilities previously split across specialized models—Magistral for logic, Pixtral for vision, and Devstral for programming—into a single, versatile powerhouse.
One of the standout features of this release is the introduction of variable reasoning effort. Users can toggle between "none" and "high" settings, allowing the model to allocate more cognitive resources to complex problems, much like the "thinking" time seen in other top-tier reasoning models. The model is released under the permissive Apache 2.0 license, weighing in at 242GB on Hugging Face, making it a substantial resource for the open-source community to host and fine-tune.
In a move catering to specialized technical niches, Mistral also introduced Leanstral. This model is specifically fine-tuned for Lean 4, a functional programming language used for formal mathematical verification and theorem proving. By targeting such a specific and rigorous domain, Mistral demonstrates the potential for smaller, specialized models to assist in high-level scientific and mathematical research where precision and formal logic are paramount.