Mistral NeMo Release Notes

Mistral AI and NVIDIA have collaborated to release Mistral NeMo, a 12B parameter model designed to provide state-of-the-art reasoning, world knowledge, and coding accuracy within its size category. This model is a drop-in replacement for systems using Mistral 7B due to its reliance on standard architecture.

Model Architecture and Performance

Mistral NeMo is a 12B parameter model that supports a large context window of up to 128k tokens. It was trained with quantization awareness, which allows for FP8 inference without any performance loss. The model is available as both pre-trained base and instruction-tuned checkpoints under the Apache 2.0 license.

Multilingual Capabilities and the Tekken Tokenizer

Mistral NeMo is optimized for global applications and is particularly strong in English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and Hindi. To achieve this, the model utilizes a new tokenizer called Tekken, based on Tiktoken.

Tekken was trained on more than 100 languages and offers superior compression efficiency compared to the SentencePiece tokenizer used in previous Mistral models. Specific compression improvements include:

  • Source code and European languages: Approximately 30% more efficient compression for source code, Chinese, Italian, French, German, Spanish, and Russian.
  • Korean and Arabic: 2x to 3x more efficient compression.
  • General Language Proficiency: Tekken is more proficient at compressing text for approximately 85% of all languages compared to the Llama 3 tokenizer.

Instruction Fine-Tuning and Alignment

Mistral NeMo has undergone an advanced fine-tuning and alignment phase, resulting in significant improvements over Mistral 7B in the following areas:

  • Instruction Following: Improved ability to follow precise instructions.
  • Reasoning: Enhanced reasoning capabilities.
  • Multi-turn Conversations: Better handling of multi-turn dialogue.
  • Code Generation: Increased accuracy in generating source code.

Availability and Deployment

Mistral NeMo is available through several channels for researchers and enterprises:

  • HuggingFace: Weights for both base and base-instruct models are hosted here.
  • Mistral AI Tools: The model can be used with mistral-inference and adapted via mistral-finetune.
  • la Plateforme: Available under the name open-mistral-nemo-2407.
  • NVIDIA NIM: The model is packaged as an NVIDIA NIM inference microservice available at ai.nvidia.com.

Sources

Related

  • Dispatch
    Pixtral 12B Release Notes

    Mistral AI has released Pixtral 12B, a natively multimodal model that combines a new 400M parameter vision encoder with a 12B parameter decoder to deliver high-performance multimodal reasoning without compromising text capabilities.

  • Dispatch
    Mistral 7B Release Notes

    Mistral AI has released Mistral 7B, a 7.3B parameter language model that outperforms Llama 2 13B across all benchmarks and is available under the Apache 2.0 license.

  • Dispatch
    Mistral Small 3.1 Release Notes

    Mistral AI has released Mistral Small 3.1, an Apache 2.0 licensed model featuring multimodal understanding, a 128k token context window, and performance that surpasses GPT-4o Mini and Gemma 3.

  • Dispatch
    Mistral Large 2 release notes / what's new

    Mistral AI has released Mistral Large 2, a 123B parameter model designed for high throughput single-node inference with competitive performance in coding, reasoning, and multilingual capabilities.

  • Dispatch
    Mistral Small 4 release notes / what's new

    Mistral AI has released Mistral Small 4, a unified open-source model under Apache 2.0 that integrates reasoning, multimodal, and agentic coding capabilities into a single versatile architecture.