LLaVA-1.5

LLaVA-1.5, released October 5, 2023 by researchers at UW-Madison and Microsoft, is the open multimodal model that defined the vision-language baseline recipe. It pairs a CLIP ViT-L/14 vision encoder with a Vicuna language model through a two-layer MLP projector, adds academic visual-QA data to the training mix, and reportedly reached state-of-the-art results across 11 multimodal benchmarks — all trained in about a day on a single 8-GPU node.

It is genuinely open under Apache 2.0, which is why it became the backbone of hundreds of derivative vision models and the foundation for the LLaVA-NeXT line. Limitations are those of its era: 336px image resolution, single-image reasoning, and far weaker OCR and instruction-following than current VLMs.