Mistral AI has unveiled a new family of open, multilingual, and multimodal models under the Mistral 3 series, ranging from compact edge-optimized variants to large-scale mixture-of-experts architectures.

The lineup includes three "Ministral" models with 3 billion, 8 billion, and 14 billion parameters, alongside the flagship "Mistral Large 3." Built on a sparse mixture-of-experts design and trained using approximately 3,000 Nvidia H200 GPUs, the flagship model features 41 billion active parameters out of a total of 675 billion.

Mistral Large 3 is fully open-source under the Apache-2.0 license. The company aims for it to rival top-tier open models in general language tasks while also supporting image understanding. On the LMArena leaderboard, it currently ranks second among open non-reasoning models and sixth among open reasoning models. Benchmark results show performance on par with other leading open models like Qwen and Deepseek—though Deepseek released version V3.2 yesterday, which demonstrates significant improvements over its predecessor across multiple evaluations.

What the New Edge Models Mean for Efficiency

The smaller "Ministral 3" variants are tailored for local and edge deployment. Available in 3B, 8B, and 14B sizes, each comes in base, "Instruct," and "Reasoning" versions—all equipped with image comprehension capabilities and released under the Apache-2.0 license.

According to Mistral, the instruction-tuned models deliver performance comparable to similar open alternatives while generating significantly fewer tokens. The reasoning variants are optimized for deeper analytical workloads, with the 14B model achieving an 85% score on the AIME-25 benchmark.

These models are accessible via Mistral AI Studio, Hugging Face, and major cloud platforms including Amazon Bedrock, Azure Foundry, IBM WatsonX, and Together AI. Support for Nvidia NIM and AWS SageMaker is also planned. Mistral noted close collaboration with Nvidia throughout the development of these new models.