Cambrian-S
Joint multimodal model from Meta/NYU/Google Brain/Stanford — spatial super-perception
A joint multimodal model from Meta, NYU, Google Brain, and Stanford with spatial super-perception.
Open foundation models — the base layers everything builds on.
Key features
- Video MLLM focused on spatial supersensing
- Predictive-sensing trained 7B model plus standard 0.5B to 7B variants
- Built on Qwen2.5-Instruct with siglip2 vision encoder
- VSI-590K spatial dataset and VSI-SUPER benchmark released
- Training code and evaluation suite open-sourced
Good for
- Spatial reasoning over video
- Embodied AI and robotics perception research
Stack & tags
PythonTorchXLAmultimodalvideo-understandingspatial-reasoningopen-weightsdatasetbenchmarkvlm