Mistral NeMo, a new AI model with 12 billion parameters and a 128K context window, is now available on MLX. Users can run inference or train the model locally on a Mac using MLX by installing the necessary packages. The model supports function calling and can be installed via 'pip install -U fastmlx' and 'mlx-lm' from source (PR #895). Additionally, Mistral NeMo is integrated with NVIDIA's NIM inference microservice, offering performance-optimized inference with TensorRT-LLM engines. The model runs at 37 tokens per second on an M3 Max and can be fine-tuned using only 12GB of VRAM, making it twice as fast and 60% more memory-efficient without accuracy loss. Fine-tuning is also supported on Google Colab using the Unsloth library. The model, including variants like Mistral NeMo Instruct (Q4), Mistral-7B, Mixtral-8x7B, and Mixtral-8x22B, is designed to meet diverse business needs and can be run with minimal code.