LLaVA-OneVision-1.5

Fully open, budget-honest LMM training: the complete framework — 85M-sample pretraining and 22M-sample instruction datasets, training code, and logs — for building competitive multimodal models for roughly $16K of compute. The 8B model beats Qwen2.5-VL-7B on 18 of 27 benchmarks. Released as 4B and 8B instruct checkpoints with all intermediate stages.

Joint LMMs-Lab/NTU (Ziwei Liu, Bo Li) and DeepGlint et al. co-development — a democratization statement in a line that began as LLaVA-OneVision with ByteDance. Model backbones initialize from open LLM/ViT components rather than a fully from-scratch pretrain.