Ministral-3

Mistral AI’s premier sub-10B parameter models engineered for high-performance edge computing and low-latency local inference.

The Ministra-3 family (comprising the 3B and 8B models) sets a new benchmark for on-device intelligence. These models support a 128k token context window and outperform Llama 3.2 3B on critical metrics: including MMLU and GSM8K. Built for efficiency: they handle complex reasoning and multilingual tasks directly on consumer hardware (like laptops and mobile devices). Deploy them for privacy-first local agents or as high-speed routers within larger AI orchestration pipelines.

https://mistral.ai/news/ministra-3b-8b/