Velvet helps train multimodal models

Velvet is a product and research company focused on multimodal models.

We work with AI labs and enterprises to create datasets & evals that advance AI's interactive capabilities.

As of June 28th, 2026, nearly all AI models lack spatial reasoning capabilities and respond with uncanny latency. We are accelerating a future in which models understand the world as we do. If you're training or deploying multimodal models, we'd love to work with you.