创始效应塑造开放LLM家族中多模态的进化动态

Large language model (LLM) families are evolving rapidly, yet the emergence and propagation speed of multimodal capabilities remain unclear. Utilizing the ModelBiome AI Ecosystem dataset, which encompasses Hugging Face model metadata and lineage fields (over 1.8 million model entries), this study quantifies multimodality over time and along recorded parent-child relationships. Cross-modal tasks are prevalent in the broader ecosystem well before they become common within major open LLM families. Within these families, multimodality remains rare through 2023 and most of 2024, then sharply increases between 2024 and 2025, predominantly driven by image-text vision-language tasks. The first vision-language model (VLM) variants typically emerge during this period across major families.