Introducing our Artifacts Hub and Adoption Dashboard

What it claims

Interconnects has launched two free, continuously updated data products on open-weights models. The Artifacts Hub covers 792 hand-selected models released in the last two years, spanning text-focused language models and multimodal generative models. The Adoption Dashboard updates daily with download and derivative-model counts by geography and organization. It replaces the US-vs-China adoption plot from The ATOM Project, which the authors say they only refreshed occasionally.

How it works

The 792 models are chosen in three stages:

1. Interconnects tracks data for every model on Hugging Face.
2. It analyzes a "core few thousand" LLMs. This list is public on GitHub and regularly updated, though the page does not give the link.
3. It hand-selects a few hundred of those for the Hub.

For each model, the Hub combines:

• Artificial Analysis Intelligence Index: shows how far behind frontier intelligence the model was.
• Hugging Face and OpenRouter adoption: downloads and inference tokens, compared with similar models.
• Relative adoption metric (RAM): downloads normalized for model size and time since release. The authors present it as part of a method for judging adoption of large MoE models, where raw download counts mislead.
• VAIL similarity index: compares a model with models from related generations.

The page shows GLM-5.2 as an example entry.

How to use it

• Need a maintained, curated list of significant LLMs? The public GitHub list of the core few thousand LLMs is the most reusable piece. It is broader than the Hub's 792.
• Need raw download comparisons? Use RAM rather than raw counts when comparing models of different sizes or release dates.
• Want the curated data for your own product or research? The authors invite this at mail@interconnects.ai. The page names no API, export format or licence, so access terms come from that contact.

Limits

• Hand-picked coverage. The Hub is curated, not exhaustive, and the page gives no inclusion criteria beyond "the open models that matter."
• RAM is underspecified. It is described only as "time-size normalized downloads." No formula, baseline or validation is given.
• Unexplained metrics. The VAIL similarity index is not explained.
• Inherited bias. All scores come from third-party sources: Hugging Face, OpenRouter and Artificial Analysis. They carry those sources' coverage gaps. For example, models not served on OpenRouter will have no inference-token data.
• Funding. The Hub was built with support from Project VAIL, an AI verification startup that also supplies one of the metrics.