Forecasting Open-Weight AI Model Growth on HuggingFace

TL;DR

The paper tackles forecasting the growth of open-weight AI models on HuggingFace by translating a scientific citation model into model adoption dynamics. It introduces three parameters—immediacy, longevity, and relative fitness—and formalizes the dynamics with a cumulative count $c_i^t$ that follows a saturating trend, e.g., $c_i^t = m ( e^{ lambda_i Phi( (ln t - mu_i)/ sigma_i ) } - 1 )$, where $m$ is a normalization and $Phi$ is the standard normal CDF. Empirical results indicate most models are well captured by the framework; some models show abrupt surges, and organization-level factors shape adoption trajectories, with Meta, BAAI, and StabilityAI showing early peaks. The framework offers a quantitative tool for forecasting long-term diffusion and informing governance and strategic decisions, with potential for extension to include additional data.

Abstract

As the open-weight AI landscape continues to proliferate-with model development, significant investment, and user interest-it becomes increasingly important to predict which models will ultimately drive innovation and shape AI ecosystems. Building on parallels with citation dynamics in scientific literature, we propose a framework to quantify how an open-weight model's influence evolves. Specifically, we adapt the model introduced by Wang et al. for scientific citations, using three key parameters-immediacy, longevity, and relative fitness-to track the cumulative number of fine-tuned models of an open-weight model. Our findings reveal that this citation-style approach can effectively capture the diverse trajectories of open-weight model adoption, with most models fitting well and outliers indicating unique patterns or abrupt jumps in usage.

Figures (11)

• Figure 1: Monthly number of fine-tuned models after a base model’s release, with colors denoting the time when it was created.
• Figure 2: (a) Distribution of values for $\lambda$, $\mu$, and $\sigma$. (b) Pairwise relationships among immediacy ($\mu_i$), longevity ($\sigma_i$), and relative fitness ($\lambda_i$) on log-scale axes.
• Figure 3: Density plots illustrating the cumulative number of fine-tuned models for relative fitness of ($1 \leq \lambda_i \leq 10$) at the 2-month, 6-month, and 12-month marks, segmented by companies. IBM and CohereAI are omitted from the 12-month plot due to the absence of models older than 12 months.
• Figure 4: Monthly cumulative number of fine-tuned models following the release of the base model, with colors indicating the base models' creation years, illustrating trends in fine-tuning patterns over time.
• Figure 5: Distribution of downloads and fine-tuned models across Hugging Face models, illustrating a strong Pareto-like concentration. A small fraction of models (1.26% for fine-tuning, 0.035% for downloads) accounts for 80% of total activity, highlighting the dominance of a few models in the open-source AI ecosystem.