Hugging Face Hub v1.0 breaks with its past, swaps backend to httpx and hf_xet

The huggingface_hub Python library just hit version 1.0, and it’s not a cosmetic update. This is a hard break designed to carry the platform through the next decade. The library now powers 200,000 dependent projects and provides the plumbing for over 2 million public models. To get there, the team, drawing from a community of nearly 300 contributors, made foundational changes that will ripple through the ecosystem.

Under the hood, the entire backend has been swapped to httpx, replacing the legacy requests-based architecture. The old huggingface-cli is dead and deprecated; a completely redesigned CLI built on Typer takes its place with a much broader feature set. Perhaps the most impactful change for anyone pushing large model files is the full adoption of hf_xet for file transfers, which replaces the aging hf_transfer module for a significant performance lift. The team at Hugging Face insists they’ve worked to maintain backward compatibility, but the seams are already showing with their own flagship libraries.

The immediate friction point is transformers. Hugging Face’s own model library will explicitly require huggingface_hub v0.x for its current v4 releases, while the upcoming transformers v5 release will lock to the new v1.x. This means anyone in the middle of production pipelines needs to be very careful about their dependency constraints. It’s a rare, explicit fork in the road between two core packages from the same organization, and it signals that v1.0 contains changes deep enough to warrant a hard version split.

Looking back, it’s a remarkable arc. The library started in late 2020 as a simple Git wrapper extracted from transformers so other projects could share model checkpoints without relying on broken Google Drive links. The real pivot came in June 2022 with version 0.8.1, which introduced the HTTP Commit API. That move was philosophical—it redefined the tool from a Git helper into a dedicated infrastructure layer for machine learning artifacts, adding a shared cache system with deduplication across all libraries. Now having absorbed full management of Spaces, Inference Endpoints, and community features like pull requests and collections, v1.0 completes that transformation. It’s no longer a utility; it’s the operating system for the Hugging Face Hub.