With a resounding announcement on 2 December 2025, French AI startup Mistral unveiled its new model family: Mistral 3. The release marks a bold step in the global race for cutting-edge, open artificial intelligence. The line includes a powerful new frontier model – Mistral Large 3 – and three smaller, highly efficient models (Ministral 3B, 8B, 14B), all multimodal, multilingual, and openly licensed under Apache 2.0. Their goal: to serve everything from enterprise-scale infrastructure to resource-constrained edge devices – and to offer a credible, independent alternative to closed ecosystems from OpenAI, Google and Anthropic.
One family, every scenario
At the heart of the release is Mistral Large 3, a sparse Mixture-of-Experts (MoE) model with 675 billion total parameters, 41 billion of which are active at any given time. Trained on around 3,000 NVIDIA H200 GPUs, the model delivers impressive performance across benchmarks: strong multilingual capabilities, robust visual understanding, and impressive results in complex reasoning tasks. It’s the most powerful model Mistral has built to date – and crucially, it’s fully open-weight.
Alongside it comes the Ministral 3 series – dense models with 3B, 8B, and 14B parameters, each available in base, instruct, and reasoning variants. These are built for maximum runtime efficiency and punch well above their weight, especially in edge environments like laptops, Jetson modules, and embedded systems. Even the smallest, 3B model runs smoothly on constrained hardware – without compromising on output quality.
Multimodal, multilingual, truly open
All Mistral 3 models are natively multimodal (text + image), support over 40 languages, and are released under the permissive Apache 2.0 licence. This grants full commercial use, fine-tuning rights, and on-prem deployment – free from restrictive terms. The models are also available in compressed formats (e.g. NVFP4) for efficient deployment on standard GPU nodes.
Hardware, partners and deployment
Mistral worked closely with NVIDIA, Red Hat and vLLM to co-design and optimise its latest generation. The models were trained using Hopper GPUs and benefit from specialised MoE kernels, speculative decoding, and long-context enhancements. Mistral 3 is already available or arriving shortly on platforms like Hugging Face, Amazon Bedrock, Azure Foundry, IBM WatsonX, Together, OpenRouter and Fireworks – with support for NVIDIA NIM and AWS SageMaker coming soon.
Positioning in the global AI landscape
With Mistral 3, the company solidifies its position as a serious contender in the AI space – not just by competing with proprietary giants, but by doubling down on transparency and accessibility. The open-weight models deliver strong performance while giving enterprises full control over deployment and customisation. Meanwhile, Mistral’s growing portfolio of domain-specific models – including Codestral, Devstral and Voxtral – demonstrates its ambition to serve a wide range of industrial and business needs.
The bigger picture
Mistral 3 isn’t just a model release – it’s a statement. It champions open, European AI innovation that rivals the best of Silicon Valley. And it does so with remarkable engineering, practical versatility, and a commitment to democratic access. Whether you’re running inference in a datacentre or deploying on a drone, Mistral 3 offers a serious, scalable and sovereign alternative in the AI arms race.
