Data Machina #260

Substack · 2 years ago · 19

Vision-language models are experiencing rapid development, with large foundation models from OpenAI, Anthropic, and Google dominating benchmarks while smaller, specialized VLMs are emerging as efficient alternatives. Five notable open or open-weight small VLMs have been released recently—LLaVA-Next, PaliGemma, Phi-3 Vision, Florence-2, and InternLM-XComposer 2.5—with several achieving state-of-the-art performance on multiple benchmarks despite their reduced size. This shift toward smaller, more efficient VLMs enables wider deployment and customization while maintaining competitive performance across vision and language tasks.