Mistral 3 launches with enterprise-grade open multimodal models, strong benchmarks, edge-optimized versions and broad cloud availability.
Mistral AI has dropped the latest version of its open multimodal and multilingual AI model: Mistral 3.
Mistral 3 includes three small, dense models (14B, 8B and 3B) along with Mistral Large 3, a mixture-of-experts (MoE) trained on 41 billion active parameters and 675 billion total parameters.
This latest model series reportedly achieves closed-source-level results while still maintaining the transparency and control of open-source models, the company asserts.
How Does Mistral 3 Compare to Other Leading Models?
According to company officials, Mistral Large 3 beats out both DeepSeek V3.1 and Kimi K2 for general prompts and multilingual prompts.
The AI company claims its new model is "one of the best permissive open weight models in the world," trained on 3,000 of NVIDIA's H200 GPUs. Mistral 3 can understand text, images and complex logic across more than 40 native languages.
On the LMArena leaderboard, Mistral Large 3 ranks at No. 2 in the OSS non-reasoning models category and No. 6 amongst OSS models overall.
Mistral worked with NVIDIA, vLLM and Red Hat to make Mistral Large 3 run faster and be more accessible for those in the open-source community. This optimized model:
Uses less memory
Runs more efficiently
Works on NVIDIA’s newest supercomputer-style systems OR
Works on a single powerful server that many companies already have
More developers can run the model themselves, instead of needing massive infrastructure.
For edge and local use cases, the Ministral 3 series is available in three model sizes: 3B, 8B and 14B parameters. Each model size offers base, instruct and reasoning variants, and are optimized to run on:
Workstations
Consumer PCs and laptops
Edge devices and robots
As a result, the same AI model family can run anywhere — from data centers to small devices.
How to Use Mistral 3
Mistral 3 is available on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face (Large 3 & Ministral), Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI and Together AI.
It will also soon be available on NVIDIA NIM and AWS SageMaker.
According to company officials, its latest offerings are ideal for enterprise use cases as well as edge-optimized solutions. Use cases include:
Coding
Creative collaboration
Document analysis
Tool-use workflows
About the Author
Michelle Hawley is an experienced journalist who specializes in reporting on the impact of technology on society. As editorial director at Simpler Media Group, she oversees the day-to-day operations of VKTR, covering the world of enterprise AI and managing a network of contributing writers. She's also the host of CMSWire's CMO Circle and co-host of CMSWire's CX Decoded. With an MFA in creative writing and background in both news and marketing, she offers unique insights on the topics of tech disruption, corporate responsibility, changing AI legislation and more. She currently resides in Pennsylvania with her husband and two dogs. Connect with Michelle Hawley:
