The smallest model in the Ministral 3 family, combining a 3.4B language model with a 0.4B vision encoder for efficient edge deployment.