Mistral Artificial Intelligence today announced the mistral 3 family of open-source multilingual, multimodal models, optimized for deployment across NVIDIA supercomputing and edge platforms. The flagship Mistral Large 3 uses a mixture-of-experts architecture that activates only the most relevant parts of the model per token to improve efficiency and scalability. The company says the family will be available everywhere, from the cloud to the data center to the edge, starting Tuesday, Dec. 2.
Mistral Large 3 is described with 41B active parameters, 675B total parameters and a large 256K context window to support enterprise artificial intelligence workloads. NVIDIA hardware and software optimizations are a central part of the announcement. By combining GB200 NVL72 systems with the model’s MoE design and tapping NVIDIA NVLink’s coherent memory domain and wide expert parallelism optimizations, the partners report material performance and cost benefits. The model also benefits from low-precision NVFP4 and NVIDIA Dynamo disaggregated inference optimizations. On the GB200 NVL72, Mistral Large 3 achieved 10x performance gain compared with the prior-generation NVIDIA H200, which the companies say translates into a better user experience, lower per-token cost and higher energy efficiency.
The release also includes nine small language models and a compact Ministral 3 suite targeted at edge devices. Those compact models are optimized for NVIDIA Spark, RTX PCs and laptops and NVIDIA Jetson devices, and developers can run them via Llama.cpp and Ollama. NVIDIA further integrates the models with open-source NeMo tools for agent lifecycle development, including Data Designer, Customizer, Guardrails and NeMo Agent Toolkit, and has optimized inference frameworks such as NVIDIA TensorRT-LLM, SGLang and vLLM for the mistral 3 family. The models are openly available on leading open-source platforms and cloud providers and are expected to be deployable soon as NVIDIA NIM microservices, offering a path from research to production for enterprise artificial intelligence use cases.
