Mistral AI and NVIDIA unveil 12B NeMo model

Mistral AI has unveiled NeMo, a powerful 12B AI model developed with NVIDIA. NeMo offers an impressive context window of up to 128,000 tokens, delivering top-tier performance in reasoning, world knowledge, and coding accuracy for its size.

The partnership with NVIDIA has produced a model that excels in performance and ease of use. NeMo is designed to seamlessly replace systems using Mistral 7B, utilizing standard architecture for easy integration.

To promote adoption and research, Mistral AI has released both pre-trained base and instruction-tuned checkpoints under the Apache 2.0 license, appealing to researchers and enterprises alike.

A standout feature of NeMo is its quantisation awareness during training, allowing for FP8 inference without performance loss. This is crucial for organizations deploying large language models efficiently.

Mistral NeMo’s performance has been compared to other models, including Gemma 2 9B and Llama 3 8B. It is designed for global, multilingual applications, with strengths in languages such as English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi.

NeMo introduces Tekken, a new tokeniser based on Tiktoken, trained on over 100 languages. Tekken improves compression efficiency for natural language text and source code by about 30% compared to previous models, with notable gains for Korean and Arabic.

Mistral NeMo's weights are available on HuggingFace for both base and instruct versions. Developers can experiment with NeMo using the mistral-inference tool and adapt it with mistral-finetune. The model is also accessible on Mistral’s platform as open-mistral-nemo.

In collaboration with NVIDIA, NeMo is available as an NVIDIA NIM inference microservice through ai.nvidia.com, streamlining deployment for NVIDIA AI ecosystem users.

The launch of Mistral NeMo marks a significant advancement in AI accessibility, combining high performance, multilingual capabilities, and open-source availability to serve a broad range of applications across various industries and research fields.