huggingface_hub v1.0: Five Years of Building the Foundation of Open Machine Learning

Source: HuggingFace Date: 2025-10-27 URL: https://huggingface.co/blog/huggingface-hub-v1

Summary

Library update: huggingface_hub reaches v1.0 after 5 years. Major changes: HTTP backend migrated from requests to httpx (HTTP/2, thread safety, unified async); file transfer switched from hf_transfer to hf_xet (64KB chunk-level deduplication, migrated 77PB across 6M repos); CLI redesigned from huggingface-cli to hf command. Breaking: Git-based Repository class removed, InferenceApi → InferenceClient, transformers v4 requires v0.x. Scale: 113.5M monthly downloads, 200K+ dependent libraries.

Implications

Thread: HF as open-source ML hub / transformers library trajectory. v1.0 of huggingface_hub is a dependency earthquake — 200K+ dependent libraries need migration, and transformers v4 is locked to v0.x while v5 requires v1.x. The hf_xet chunk-level deduplication migration (77PB) is a backend infrastructure achievement that improves transfer efficiency for the entire HF ecosystem silently. The HTTP/2 + httpx migration enables proper async from the SDK, which is important for any application making concurrent Hub requests. Watch the transformers v5 release timeline — that’s the coupling point for when v1.0 adoption becomes the mainstream path.