Mistral launched its latest generation of artificial intelligence (AI) models on Tuesday, the Mistral 3 family that includes both powerful large-scale systems and compact edge-optimized models.
The Microsoft Corp.-backed company’s flagship release, Mistral Large 3, is a sparse mixture-of-experts architecture with 41 billion active parameters and 675 billion total parameters on 3,000 of NVIDIA Corp.’s H200 GPUs. According to Mistral, this makes Large 3 “one of the best permissive open weight models in the world.”
The model debuted at No. 2 in the open-source non-reasoning models category on the LMArena leaderboard, placing sixth overall among all open-source models. After post-training refinement, Mistral Large 3 achieved performance parity with leading instruction-tuned open-weight models while excelling in image understanding and multilingual conversations, particularly in languages beyond English and Chinese, Mistral said.
Rather than activating every component for each task, Mistral Large 3’s mixture-of-experts design selectively engages only the most relevant portions of the model. Its granular routing leverages NVIDIA’s NVLink coherent memory and parallelism capabilities, along with the company’s NVFP4 low-precision format that preserves accuracy. Additional performance optimizations include NVIDIA Dynamo, TensorRT-LLM, SGLang, and vLLM.
The Mistral 3 family extends beyond large-scale models. The company released three smaller dense models at 14 billion, 8 billion, and 3 billion parameters, plus nine compact Ministral 3 models specifically optimized for NVIDIA’s edge platforms, including Spark, RTX PCs and laptops, and Jetson devices.
All Mistral 3 models were trained on NVIDIA Hopper GPUs, using high-bandwidth HBM3e memory for frontier-scale workloads. The models became available Tuesday across multiple platforms, including Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI. NVIDIA NIM microservices and AWS SageMaker support is expected soon.
The announcement comes as competition intensifies in the AI sector. Earlier Tuesday, reports emerged that Microsoft-backed OpenAI is developing a large language model called Garlic to counter Google’s recent advances in AI development.
Mistral’s momentum continues from September when the startup reportedly entered final negotiations to raise $2.3 billion at a $13.9 billion valuation, underscoring investor confidence in the French company’s position in the AI market.
