cambrian-s

Why we included this project

Cambrian-S is a family of vision-language models built for spatial reasoning in video, an area where general multimodal LLMs still tend to fall short. The repo ships checkpoints from 0.5B up to 7B parameters along with the training code and an evaluation suite, so you can run the pretrained models directly or reproduce and extend the training recipe yourself. It also publishes the VSI-590K instruction-tuning dataset and the VSI-SUPER benchmark, which are handy if you want to fine-tune your own spatial video model or score it against a dedicated test. For teams building video understanding or embodied perception pipelines, the weights and evaluation harness are ready to adapt; researchers get a training stack they can reproduce and build on.