Mistral AI and NVIDIA have released Mistral NeMo 12B, a new language model aimed at enterprises. The model excels in multi-turn conversations, math, common sense reasoning, world knowledge, and coding. Mistral touts its unprecedented accuracy and flexibility for diverse enterprise applications.
Mistral NeMo has a 128K context length, allowing it to handle complex, lengthy inputs more effectively than many competitors. The model processes extensive information coherently, producing contextually relevant outputs for diverse enterprise needs.
Guillaume Lample, Mistral AI's cofounder and chief scientist, highlighted the collaboration's benefits: "Leveraging NVIDIA's hardware and software let us develop a model with high accuracy, flexibility, and efficiency, backed by enterprise-grade support and security."
The model's training utilized NVIDIA's AI infrastructure, including 3,072 H100 80GB Tensor Core GPUs on the DGX Cloud AI platform. This process incorporated accelerated training techniques to optimize performance.
Mistral NeMo offers several key advantages for enterprise users:
- Versatility: Released under the Apache 2.0 license, it's designed as a drop-in replacement for systems using Mistral 7B.
- Efficiency: Uses FP8 data format for inference, reducing memory needs and speeding up deployment without accuracy loss.
- Multilingual capability: Excels in English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi.
- Improved tokenization: Introduces Tekken, a new tokenizer based on Tiktoken, showing efficiency gains across 100+ languages.
- Easy deployment: Packaged as an NVIDIA NIM inference microservice, allowing rapid setup in various environments.
- Hardware flexibility: Runs on a single NVIDIA L40S, GeForce RTX 4090, or RTX 4500 GPU, balancing performance and cost.
For businesses looking to implement advanced AI capabilities, Mistral NeMo 12B offers a potent combination of features. Its multilingual proficiency and efficient processing make it suitable for a wide range of enterprise applications.
Moreover, Mistral NeMo is designed with enterprise-grade security and support, featuring dedicated feature branches, rigorous validation processes, and comprehensive service-level agreements. Enterprises can seamlessly integrate Mistral NeMo into their commercial applications, benefiting from direct access to NVIDIA AI experts and reliable, consistent performance.
Mistral NeMo is available now through ai.nvidia.com, with a downloadable NIM version coming soon.