Light-Omni

Light-Omni is a multimodal agent framework for reflexive video understanding with long-term memory. It replaces costly detective-style iterative reasoning with dual contextual states: a compact global state consolidated from episodic memory, and a latent state that drives action control and semantically aligned retrieval.

This repository hosts the Light-Omni model checkpoint for inference. It contains the safetensors weight shards, tokenizer files, model configuration, and multimodal preprocessor configuration files.

Links

• Project page: https://clare-nie.github.io/Light-Omni/
• Code: https://github.com/Clare-Nie/Light-Omni
• Dataset: https://huggingface.co/datasets/ClareNie/Light-Omni-Training
• Model: https://huggingface.co/ClareNie/Light-Omni
• Paper: https://arxiv.org/abs/2607.05511

Citation

@misc{nie2026lightomni,
  title={Light-Omni: Reflex over Reasoning in Agentic Video Understanding with Long-Term Memory},
  author={Nie, Chang and Wei, Jiaju and Feng, Junlan and Fu, Chaoyou and Shan,
  Caifeng},
  year={2026},
  eprint={2607.05511},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={http://arxiv.org/abs/2607.05511}
}