Study reveals multimodal AI capabilities lag significantly behind text-only models in open LLM families, with vision-language variants appearing months to years after initial releases.

Analysis of 1.8 million+ Hugging Face models shows multimodality remained rare through 2023 and most of 2024 before sharply increasing in late 2024-2025.

Vision-language model variants lag text-generation releases by 1 month (Gemma) to over 26 months (GLM) across major open LLM families.

Image-text vision-language tasks now dominate multimodal capabilities in open LLM ecosystems, despite cross-modal tasks being widespread earlier in the broader AI ecosystem.

The ModelBiome dataset tracked parent-to-child model lineages to quantify how multimodal features propagate through LLM family trees.