ION expands ETF trading capabilities with Tradeweb integration
Investing.com -- Mistral AI announced the release of Mistral 3, its next generation of open multimodal and multilingual AI models on Tuesday.
The new lineup includes three small, dense models (14B, 8B, and 3B parameters) called Ministral, and Mistral Large 3, the company’s most capable model to date featuring a sparse mixture-of-experts architecture with 41B active and 675B total parameters.
All models are released under the Apache 2.0 license, allowing developers to freely use and modify them. The Ministral models are designed for edge and local use cases, while Mistral Large 3 is positioned as one of the best permissive open weight models globally.
Mistral Large 3 was trained from scratch on 3,000 of NVIDIA’s H200 GPUs and debuts at #2 in the OSS non-reasoning models category on the LMArena leaderboard. The company is releasing both base and instruction fine-tuned versions, with a reasoning version coming soon.
The company collaborated with NVIDIA, vLLM, and Red Hat to optimize the models. Mistral 3 models were trained on NVIDIA Hopper GPUs to utilize high-bandwidth HBM3e memory for large-scale workloads.
For edge computing, the Ministral 3 series comes in three sizes (3B, 8B, and 14B parameters), each with base, instruct, and reasoning variants that include image understanding capabilities. According to Mistral AI, these models achieve the best cost-to-performance ratio of any open-source model.
Mistral 3 is immediately available on multiple platforms including Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI. The models will soon be available on NVIDIA NIM and AWS SageMaker.
For organizations seeking customized solutions, Mistral AI also offers custom model training services to fine-tune or adapt their models to specific needs.
This article was generated with the support of AI and reviewed by an editor. For more information see our T&C.