Technical Specifications

System Requirements

VRAM requirements for different quantization methods and context sizes

Ministral-3B-2410

Ministral-3B-2410 is a foundational language model developed by Mistral AI, specifically optimized for on-device and edge computing applications. This model is part of the 'les Ministraux' family, designed to provide computationally efficient and low-latency solutions for scenarios demanding local, privacy-first inference. Its compact size enables deployment in resource-constrained environments, including smartphones, tablets, and IoT devices. Ministral-3B-2410 can also function as an intermediary in multi-step agentic workflows, handling tasks such as input parsing, task routing, and API calls, thereby reducing latency and cost when integrated with larger models like Mistral Large.

Architecturally, Ministral-3B-2410 is a dense Transformer model. It integrates advanced attention mechanisms, including Grouped Query Attention (GQA), to enhance processing speed and manage memory overhead. The model supports a context length of up to 128,000 tokens, facilitating the processing of extended inputs for complex tasks. Consistent with other models in the Mistral AI family, it employs Rotary Position Embedding (RoPE) and RMS Normalization. The model utilizes a V3-Tekken tokenizer with a vocabulary size of 131,072.

Ministral-3B-2410 is engineered for a variety of use cases requiring local inference, such as on-device translation, internet-less smart assistants, local analytics, and autonomous robotics. It supports native function calling capabilities, making it effective for AI agents and specialized tasks. The model is designed for a balance between power efficiency and performance, leveraging pruning and quantization techniques to minimize computational load for deployment on devices with limited hardware capacity.

No evaluation benchmarks for Ministral-3B-2410 available.