Technical Specifications

System Requirements

VRAM requirements for different quantization methods and context sizes

Ministral 3 14B

Ministral 3 14B is a sophisticated, dense model within the Ministral 3 family, designed to deliver advanced capabilities while accommodating practical hardware constraints, making it suitable for a range of private AI deployments. It integrates a 13.5 billion parameter language model with a 0.4 billion parameter vision encoder, enabling multimodal understanding and processing of both textual and visual inputs. This model variant is distinguished by its multilingual capabilities, supporting over 40 languages, including major European and East Asian languages.

The architectural design of Ministral 3 14B focuses on efficient performance for edge and local computing environments. It is structured as a dense model, contrasting with Mixture-of-Experts (MoE) architectures found in larger models. The model supports a substantial context window of 256,000 tokens, facilitating the processing of extensive inputs and enabling more comprehensive interactions. Its attention mechanism is a Multi-Head Attention (MHA) structure, a standard component in transformer-based models that allows for processing different parts of the input sequence concurrently.

In terms of functionality, Ministral 3 14B offers advanced agentic features, including native function calling and structured JSON output, which enhances its utility in complex automation and conversational AI systems. This model is optimized for diverse use cases such as private chat applications, local AI assistants, and fine-tuning for specialized tasks. Its design prioritizes performance at a smaller scale, ensuring deployability across various hardware configurations, including those with limited resources.

No evaluation benchmarks for Ministral 3 14B available.