DuckDB 2.0 Ships Native hf:// Extension with Xet Dedup Saving 80% Storage

lhoestq · x · 2026-10-22

Hugging Face engineer Quentin Lhoest released a native C++ DuckDB 2.0 extension for reading and writing hf:// URLs. Reads go through the Hub's /resolve/ endpoints with range requests and a shared read-ahead cache; writes use the Hub commit API with Xet-based deduplicated uploads. HF says its Xet-backed Buckets save 80% of storage globally. Install via INSTALL hf FROM hfrepo; LOAD hf FROM hfrepo; — available at hf.co/duckdb-extensions.

Related event: Hugging Face Ships DuckDB 2.0 Extension for SQL Access to Datasets(3 posts)→

Original post →

More from Infra

Infra channel →