Cambrian-1

Saining Xie's vision-centric open multimodal LLM family (8B/13B/34B), released with the Cambrian-10M instruction dataset and the CV-Bench vision-centric benchmark (NeurIPS 2024 oral, ~2.0K★). Cambrian-1 rebalanced multimodal LLM design around visual grounding rather than language priors and anchors NYU's most consequential vision-LLM line, continued in Cambrian-S and Cambrian-P. The multimodal models are instruction-tuned on open LLM backbones rather than pretrained from scratch.