LLaVA-Video
Open 7B/72B video-language models with a 178K video instruction dataset
About
LLaVA-Video is the video branch of the LLaVA open multimodal family, released October 2024 in 7B and 72B sizes. It pairs the models with LLaVA-Video-178K, a synthetic video instruction-tuning dataset containing about 178K captions, 960K open-ended QA pairs, and 196K multiple-choice items. The models target video benchmarks including Video-MME, LongVideoBench, and Dream-1K, where the team reports competitive performance against larger systems. The paper is on arXiv as 2410.02713, and checkpoints, data, and training docs are public through the lmms-lab Hugging Face collection and the LLaVA-NeXT GitHub repo. As an academic release it is fully open for research and commercial experimentation, though production serving and long-video efficiency are community-supported. Later LLaVA work added critic-style reasoning variants, but LLaVA-Video remains the reference open model for video question answering from that generation.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: competitive on Video-MME, LongVideoBench, Dream-1K
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: Release 2024-10-04 verified via LLaVA-NeXT GitHub; 7B/72B sizes and 178K dataset confirmed; Benchmark claims team-reported.