LLaVA-OneVision
Single open model for single-image, multi-image, and video understanding
About
LLaVA-OneVision, published October 2024, was the first single open model to push the state of the art across single-image, multi-image, and video scenarios at once, with 0.5B, 7B, and 72B sizes. Its design transfers visual task ability across modalities, so image training improves video understanding. The team reports the 7B beating GPT-4V on MathVista (67.5 vs 49.9), DocVQA (91.3), and MMBench (85.9). Weights, data recipes, and training code are open, and a later 1.5 release made the pipeline fully reproducible on a reported $16,000 budget. As an academic release, production hardening and serving tooling lag the commercial labs.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
MathVista 67.5, DocVQA 91.3, MMBench 85.9, ChartQA 83.7 (team-reported, vs GPT-4V)
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: Paper submitted Oct 2024; 0.5B/7B/72B sizes and single/multi-image/video coverage verified via alphaXiv and project summary; MathVista 67.5, DocVQA 91.3, MMBench 85.9 vs GPT-4V are team-reported; labeled as such; Academic release; production tooling lags commercial labs.