Today, Mistral AI unveiled the Mistral 3 family, a new wave of open-source multilingual and multimodal models that’s been tuned to run smoothly across NVIDIA’s supercomputing platforms all the way down to edge devices.
The spotlight is definitely on Mistral Large 3, an MoE, or mixture-of-experts model, which basically means it only turns on the most relevant parts of the model for each token. Instead of wasting compute, it focuses its power where it matters, giving enterprises even more efficiency and accuracy than before – all without the usual cost and complexity. The company says it will roll out across cloud, data centers, and edge platforms starting December 2.
This powerful model will be running 41B active parameters at a given time, but it actually holds 675B in total, as well as a huge 256K context window. Paired with NVIDIA’s GB200 NVL72 systems, Mistral Large 3 taps into advanced parallelism and NVIDIA’s NVLink architecture to unlock high-performance expert parallelism. And when you layer in NVIDIA’s NVFP4 precision format and its Dynamo disaggregated inference stack, the model keeps high accuracy while still running more efficiently. On the GB200 NVL72 specifically, Mistral Large 3 hits a 10x performance jump over the previous-gen H200 GPU, which translates into faster responses, lower per-token cost, and better energy savings.
Of course, not everything is GB200 NVL72-like, so there are 9 new lightweight versions of Ministral 3 models designed for edge devices, from NVIDIA Spark and RTX PCs to Jetson-based platforms. NVIDIA’s been working closely with frameworks like Llama.cpp and Ollama to make sure these smaller models run fast on consumer GPUs, and anyone can try them out today through both tools.
Everything in the Mistral 3 suite is openly available, so developers and researchers can tinker, customize, and scale their own AI systems. And because the models plug right into open-source NVIDIA NeMo tools like Guardrails, Customizer, Data Designer and the NeMo Agent Toolkit, enterprises can take these models from prototype to production much more quickly. NVIDIA has also optimized inference stacks such as TensorRT-LLM, SGLan,g and vLLM to make sure Mistral 3 runs efficiently from cloud to edge.
The full Mistral 3 family is available now through major open-source platforms and cloud providers, with NVIDIA NIM microservice support coming soon.
