huggingfaceR 2.3.0

Optional local models

huggingfaceR 2.2.0

This release re-aligns the package with the current Hugging Face Inference Providers platform. Since mid-2025 the first-party hf-inference provider has narrowed to CPU-friendly classic models, and several models the package shipped as defaults are no longer served by it.

Scope of task-based inference

huggingfaceR’s task functions (hf_classify(), hf_summarize(), hf_embed(), hf_text_to_image() and friends) speak the Hugging Face task API contract: POST .../models/{model} with an {"inputs": ...} body. Only the first-party hf-inference provider implements that contract. Other Inference Providers (Fal AI, nscale, DeepInfra, Together, and so on) expose their own routes, payloads, and response shapes, and are reachable through the Hugging Face clients rather than this contract.

This release makes that boundary explicit instead of silently producing confusing failures. Chat (hf_chat(), hf_generate(), hf_describe_image()) is unaffected: it uses the OpenAI-compatible router endpoint, which selects a provider server-side and supports every routed chat model.

Bug fixes

New features

Improvements

Known limitations

huggingfaceR 2.1.0

New features

Improvements

huggingfaceR 2.0.0

Breaking changes

New features

Improvements