Ministral 2410 3B

Ministral-3B-Instruct-2410 is the smallest member of Mistral AI’s les Ministraux family—state-of-the-art models engineered for intelligence at the edge. With just 3 billion parameters, Ministral-3B delivers a surprising level of performance in commonsense reasoning, multilingual understanding, and task-following capabilities, while maintaining an exceptionally small memory and compute footprint. Ideal for applications where compute constraints are tight and privacy is paramount, Ministral-3B supports a 128k token context window (32k on vLLM) and shares the same tokenizer (V3-Tekken, 131k vocab) as its larger sibling, enabling consistent input handling across the stack. It is well-suited for use in smart assistants, IoT systems, on-device analytics, and mobile applications where responsiveness, cost, and local inference matter most. Though compact, Ministral-3B can be integrated into larger multi-agent architectures, acting as a lightweight parser or router in workflows coordinated by models like Mistral Large