Mistral AI has unveiled its Mistral 3 family of open-source multilingual, multimodal models optimised for NVIDIA’s supercomputing and edge platforms from data centres to RTX PCs and Jetson devices.
The full Mistral 3 lineup — from flagship MoE Mistral Large 3 to compact Ministral 3 suite — was trained on NVIDIA GPUs and optimised for peak performance across NVIDIA hardware.
The Mistral Large 3 flagship is a mixture-of-experts (MoE) model with 41 billion active parameters and 675 billion total. It features a 256K context window for scalable enterprise workloads such as document analysis and agentic tasks.
NVIDIA’s optimisations deliver 10x performance gains on GB200 NVL72 systems over prior H200 GPUs, leveraging NVLink, NVFP4 low-precision inference, and frameworks like TensorRT-LLM and vLLM.
The compact Ministral 3 series — 3B, 8B and 14B variants — offers top cost-performance for edge AI, supporting image understanding and multilingual tasks under Apache 2.0 licence.
All models are available immediately on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face (Large 3 & Ministral), Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI, with NVIDIA NIM microservices forthcoming for production deployment.
Entelechy Asia 