VideoChat-Flash

VideoChat-Flash (arXiv 2501.00574, December 2024; accepted to ICLR 2026) is Shanghai AI Laboratory's answer to the context problem in video LLMs. Its core is HiCo, hierarchical visual-token compression that exploits redundancy across clips to shrink long video context from clip level to video level — roughly a 1/50 compression ratio with, the authors claim, almost no performance loss. Around that sit a multi-stage short-to-long training scheme, the LongVid dataset of real-world long videos, and a multi-hop needle-in-a-video-haystack benchmark. The reported results are striking: 99.1% accuracy over 10,000 frames in NIAH (claimed as the first open-source model to do it), plus leading 2B/7B results on long and short video benchmarks that the authors say beat much larger models including InternVL2-76B, GPT-4o and Gemini-1.5-Pro. All of that is author-reported. The repo (OpenGVLab/VideoChat-Flash) is MIT licensed with public weights, and the same HiCo idea later powered InternVideo2.5. No hosted API.

• Needle-in-a-video-haystack: 99.1% accuracy over 10,000 frames (author-reported; claimed first among open-source models)\n- Reported leading 2B/7B results on long and short video QA benchmarks, ahead of InternVL2-76B, GPT-4o and Gemini-1.5-Pro on the authors' evals (provider-claimed)\n- No independent replication of the headline figures found