Mistral drops Mistral 3 family under Apache 2.0

Mistral 3 is a comprehensive family of open multimodal models spanning lightweight edge devices to heavy-duty datacenter scale, and best of all, every single model is released under the truly permissive Apache 2.0 license with weights freely downloadable.

The headline act is Mistral Large 3, a massive sparse mixture-of-experts (MoE) model packing 675B total parameters and activating 41B per token. It handles a 256,000-token context window, natively understands images, and supports conversational tasks across more than 40 languages. On the LMArena leaderboard, Large 3 debuted at number two in the open-source non-reasoning category and number six among all open-source models.

For local and edge builds, Mistral dropped the Ministral 3 lineup. It comes in 3B, 8B, and 14B sizes, with each size offering Base, Instruct, and Reasoning variants for a total of nine models, all with native vision support. Mistral notes that the 14B Reasoning variant puts up an 85% score on AIME '25.

Hardware specs and running it locally

Hardware requirements split sharply between the edge models and the flagship MoE.

Mistral Large 3 benefits from day-one optimizations across NVIDIA, vLLM, and Red Hat. Mistral released an NVFP4 quantized checkpoint that lets you serve this 675B MoE on a single node of 8x A100 or 8x H100 GPUs using vLLM, as well as on Blackwell GB200 NVL72 racks where it delivers a 10x performance leap over H200 setups.

If you want to run purely on local machines, Ministral 3 is built for you. All nine dense models run on a single GPU, and can be deployed on RTX laptops, NVIDIA Jetson modules, or DGX Spark edge setups completely offline.

You can grab the weights directly on Hugging Face right now. Edge runners can spin up Ministral models using Llama.cpp and Ollama, while cloud endpoints are already live on Mistral AI Studio, OpenRouter, Together AI, Fireworks, Unsloth AI, Azure Foundry, and Amazon Bedrock.

Why it matters

While Meta attached user-scale restrictions to Llama, Mistral gave the open-source community a genuine Apache 2.0 license on a frontier 675B model and nine edge specialists. That gives enterprise builders absolute freedom to adapt, fine-tune, and commercialize without legal landmines.

We think the edge lineup is the real sleeper hit here. Most business workflows do not need an overpriced API token meter spinning in the cloud; an 8B or 14B reasoning model running on single-GPU hardware with native vision often solves the task faster, cheaper, and entirely within private boundaries. It is worth testing right now if you build local agents or private multimodal pipelines.