# Ministral 3 Ministral 3 refers to the "Ministraux" family of small dense models released by [[Mistral AI]] in October 2024, on the first anniversary of Mistral 7B. The family includes Ministral 3B and Ministral 8B, designed for on-device and edge computing use cases. Both models support up to 128k context length (32k on vLLM at launch). Ministral 8B uses an interleaved sliding-window attention pattern for faster and more memory-efficient inference. Key characteristics: - **Ministral 3B**: smallest model, outperforms the much larger Mistral 7B on most benchmarks. Priced at $0.04/M tokens - **Ministral 8B**: uses sliding-window attention for efficient inference. Priced at $0.1/M tokens. Weights available under Mistral Research License - **Edge-first**: built for local, privacy-first inference (on-device translation, internet-less assistants, local analytics, autonomous robotics) - **Agentic workflows**: efficient intermediaries for function calling in multi-step workflows when used alongside larger models - **Strong benchmarks**: outperform peers (Gemma 2 2B, Llama 3.2 3B, Llama 3.1 8B) across knowledge, commonsense, reasoning, and function calling Ministral 8B Instruct weights are available on HuggingFace for research use. Ministral 3B is available under a Mistral Commercial License. ## References - https://mistral.ai/news/ministraux - https://huggingface.co/mistralai/Ministral-8B-Instruct-2410 ## Related - [[Mistral AI]] - [[Large Language Models (LLMs)]]