Mistral Small (latest)

Mistral Small 4 is a sparse mixture-of-experts model designed to deliver strong performance without proportional compute cost. With 119 billion total parameters that activate only 6 billion per token during inference, it balances scale and efficiency in a way that supports fast text responses alongside logical reasoning and image understanding in a single multimodal package. This architecture places it in an interesting position between compact models that sacrifice capability for speed and larger dense models that demand more infrastructure. The hybrid design unifying instruct, reasoning, and coding capabilities suggests it was built to handle the breadth of everyday AI tasks rather than excelling at a single narrow benchmark.

The model arrives as the latest iteration in Mistral's Small family, building on the momentum of Mistral Small 3's Apache 2.0 release. Early variants in this line prioritized latency and efficiency, achieving competitive results against much larger models while maintaining throughput suitable for high-volume workloads. Mistral Small 4 inherits this lineage while adding multimodal support and expanded context handling for longer documents and conversations. The Apache licensing carried forward from earlier releases makes it practical for enterprises to deploy, customize, and run at scale without licensing friction. This combination of architectural efficiency, broad capability coverage, and open deployment options positions it as a practical choice for cost-sensitive automations and background processing where reliability and speed matter more than frontier-level scale.