LongVA

LongVA is an academic open model from the EvolvingLMMs-Lab team that pushes video understanding into long-context territory. Its core idea is long-context transfer: language-model techniques for handling extended sequences are applied directly to vision, letting the model ingest thousands of video frames in a single pass instead of sampling a handful.

The 7B model is best known for long-video retrieval, with the authors reporting near-99 percent accuracy on video needle-in-a-haystack tests across 3,000 frames. It is a research artifact rather than a product: weights and code are public, there is no hosted API, and maintenance follows an academic release cadence.

Authors report ~99% accuracy on video needle-in-a-haystack over 3,000 frames (paper claim); third-party reproductions (Video-XL paper) cite LongVA-7B around 67 on VideoMME and 56 on LongVideoBench; composite figure 56.3/67.4 reported in Video-XL comparison table