A “backbone” model is designed to extract task-agnostic representations from specific data modalities (e.g., images, text, speech). This representation can then be fine-tuned for specialized tasks.

A “foundation” model is versatile and designed for multi-task capabilities, without the need for fine-tuning.

Models capable of generating text, images, or other data using generative models, often in response to prompts.

Large language models. Useful for a variety of tasks including language generation, optical character recognition, information retrieval, and more.

Mixture-of-experts models. Route each token through a small subset of expert sub-networks, delivering large model capacity at a fraction of the per-token compute.

A “real-time” model can typically achieve 5-60 predictions per second. This translates to latency ranging up to 200 ms per prediction.

Models for robotic perception, planning, and control — including vision-language-action policies. Useful for tasks like manipulation, navigation, and embodied reasoning across diverse robot platforms.

Vision language models. Understand and reason about images alongside text, useful for tasks like visual question answering, image captioning, and multimodal retrieval.