Mistral AI and NVIDIA today released a new state-of-the-art language model, Mistral NeMo 12B, that developers can easily customize and deploy for enterprise applications supporting chatbots, multilingual tasks, coding and summarization.

By combining Mistral AI’s expertise in training data with NVIDIA’s optimized hardware and software ecosystem, the Mistral NeMo model offers high performance for diverse applications.

“We are fortunate to collaborate with the NVIDIA team, leveraging their top-tier hardware and software,” said Guillaume Lample, cofounder and chief scientist of Mistral AI. “Together, we have developed a model with unprecedented accuracy, flexibility, high-efficiency and enterprise-grade support and security thanks to NVIDIA AI Enterprise deployment.”

Mistral NeMo was trained on the NVIDIA DGX Cloud AI platform, which offers dedicated, scalable access to the latest NVIDIA architecture.

NVIDIA TensorRT-LLM for accelerated inference performance on large language models and the NVIDIA NeMo development platform for building custom generative AI models were also used to advance and optimize the process.

This collaboration underscores NVIDIA’s commitment to supporting the model-builder ecosystem.

Delivering Unprecedented Accuracy, Flexibility and Efficiency 

Excelling in multi-turn conversations, math, common sense reasoning, world knowledge and coding, this enterprise-grade AI model delivers precise, reliable performance across diverse tasks.

With a 128K context length, Mistral NeMo processes extensive and complex information more coherently and accurately, ensuring contextually relevant outputs.

Released under the Apache 2.0 license, which fosters innovation and supports the broader AI community, Mistral NeMo is a 12-billion-parameter model. Additionally, the model uses the FP8 data format for model inference, which reduces memory size and speeds deployment without any degradation to accuracy.

That means the model learns tasks better and handles diverse scenarios more effectively, making it ideal for enterprise use cases.

Mistral NeMo comes packaged as an NVIDIA NIM inference microservice, offering performance-optimized inference with NVIDIA TensorRT-LLM engines.

This containerized format allows for easy deployment anywhere, providing enhanced flexibility for various applications.

As a result, models can be deployed anywhere in minutes, rather than several days.

NIM features enterprise-grade software that’s part of NVIDIA AI Enterprise, with dedicated feature branches, rigorous validation processes, and enterprise-grade security and support.

It includes comprehensive support, direct access to an NVIDIA AI expert and defined service-level agreements, delivering reliable and consistent performance.

The open model license allows enterprises to integrate Mistral NeMo into commercial applications seamlessly.

Designed to fit on the memory of a single NVIDIA L40S, NVIDIA GeForce RTX 4090 or NVIDIA RTX 4500 GPU, the Mistral NeMo NIM offers high efficiency, low compute cost, and enhanced security and privacy.

Advanced Model Development and Customization 

The combined expertise of Mistral AI and NVIDIA engineers has optimized training and inference for Mistral NeMo.

Trained with Mistral AI’s expertise, especially on multilinguality, code and multi-turn content, the model benefits from accelerated training on NVIDIA’s full stack.

It’s designed for optimal performance, utilizing efficient model parallelism techniques, scalability and mixed precision with Megatron-LM.

The model was trained using Megatron-LM, part of NVIDIA NeMo, with 3,072 H100 80GB Tensor Core GPUs on DGX Cloud, composed of NVIDIA AI architecture, including accelerated computing, network fabric and software to increase training efficiency.

Availability and Deployment

With the flexibility to run anywhere — cloud, data center or RTX workstation — Mistral NeMo is ready to revolutionize AI applications across various platforms.

Experience Mistral NeMo as an NVIDIA NIM today via ai.nvidia.com, with a downloadable NIM coming soon.