MAV (Model Activity Visualiser) provides real-time visualization of LLM internals including attention patterns, token probabilities, and layer activations. It runs as a CLI application and integrates with Hugging Face transformers models, allowing developers and researchers to inspect model behavior during inference. The tool supports multiple open-source models like GPT-2, SmolLM, and Llama, displaying activation data in an interactive visual format to aid interpretability research and debugging.