Mistral AI unveils new edge model series 'Ministral'

Marking the one-year anniversary of Mistral 7B's release, Mistral AI is introducing its next-generation models for on-device computing and edge environments: Ministral 3B and Ministral 8B. These models set a new standard for knowledge, commonsense, reasoning, function-calling, and efficiency at the sub-10B parameter scale.

Both models support context lengths of up to 128k (currently 32k on vLLM), and in particular, Ministral 8B applies a special interleaved sliding-window attention pattern for faster and more memory-efficient inference.

Key use cases include:

• Privacy-focused local inference such as on-device translation, offline smart assistants, local analytics, and autonomous robotics.
• Serving as efficient intermediaries that perform input parsing, task routing, and API calls within agentic workflows when combined with larger models like Mistral Large.

Benchmark results show that the Ministral models demonstrate superior performance across a variety of tasks compared to similarly sized models such as Gemma 2 2B, Llama 3.2 3B, and Llama 3.1 8B.