Cambrian-1

Cambrian-1 is a family of multimodal models from NYU (mid-2024) built around a vision-centric thesis: vision encoder design choices matter more than usually credited. It introduces the Spatial Vision Aggregator, a connector that fuses high-resolution vision features with the language model while cutting token counts, plus a new vision-centric benchmark, CV-Bench. The release is unusually complete for academia: weights, code, datasets, and full instruction-tuning recipes are all public, which makes it more of a research cookbook than a product. The license terms need confirmation on the model card before commercial use, and production serving support is community-driven.