Blog

Research from the BigHugger index: what gets built, what gets licensed as what, and what the numbers say when you can query every model, paper and skill at once.

1. Which models can you actually run?recomputed nightly

   Deployability by runtime, format, quantisation and licence — the questions every hosted-model leaderboard leaves out, with Rust runtimes first because nobody writes one on a model card.

   live analysis
2. What you're allowed to shiprecomputed nightly

   Commercial use, attribution and share-alike resolved for 185,014 open models, including the one in five whose terms cannot be determined at all.

   live analysis
3. The agentic layer, by structurerecomputed nightly

   Skills and MCP servers ranked by structure rather than popularity: the names everyone writes independently, the tool names converging on a standard, the repositories shipping the most.

   live analysis

1. 97% of AI papers never get an implementation2026-09-12

   Of the 77,686 papers published since 2023, the share with an implementation inside 90 days rose from 0.94% to a peak of 3.34%. Almost nothing else is built.

   research