Mistral said on 18 July 2024 that, with NVIDIA, it had released Mistral NeMo, a 12-billion-parameter model with a context window of up to 128k tokens.
Base and instruction-tuned checkpoints are under Apache 2.0. The model uses standard architecture as a drop-in replacement for systems built on Mistral 7B, and was trained with quantisation awareness so FP8 inference incurs no performance loss.
Mistral said reasoning, world knowledge and coding accuracy are state of the art in its size category versus peers such as Gemma 2 9B and Llama 3 8B. Multilingual strength spans English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic and Hindi.
A new Tekken tokenizer, based on Tiktoken and trained on more than 100 languages, is about 30% more efficient at compressing source code and several European and East Asian languages than prior SentencePiece tokenizers, and up to 3× more efficient on Arabic.
The model is on Hugging Face, exposed on la Plateforme as open-mistral-nemo-2407, and packaged as an NVIDIA NIM inference microservice on ai.nvidia.com.