Mistral Nemo Instruct 2407 is an instruct fine-tuned version of Mistral Nemo Base 2407, developed jointly by Mistral AI and NVIDIA and released under the Apache 2 license. It is designed as a drop-in replacement for Mistral 7B, with a transformer architecture comprising 40 layers, a hidden dimension of 5,120, a SwiGLU activation function, 32 attention heads paired with 8 key-value heads under grouped-query attention, and a vocabulary of roughly 128k tokens using rotary embeddings scaled to theta of 1M. The model was trained with a 128k context window on a substantial share of multilingual and code data, aiming to push performance beyond earlier models of comparable size while remaining open-weight for research and product use.

On common reasoning and knowledge benchmarks the model reports competitive results for its class, including 68.0% on MMLU 5-shot, 83.5% on HellaSwag 0-shot, and 76.8% on Winogrande 0-shot, alongside balanced multilingual MMLU scores across French, German, Spanish, Italian, Portuguese, Russian, and Chinese in the low-to-mid 60s and 59% range for Japanese. These figures suggest practical strength for conversational assistants, multilingual chat, and code-aware workflows that benefit from a very long context. The release is suited to teams that want an open multilingual model with broad tokenizer coverage, long-context handling, and flexible deployment through frameworks such as mistral-inference or transformers.