Huggingface Hub

Hugging Face hosts the world's largest model repository (150,000+ models) and provides serverless inference for most of them via the Inference API — run a text classification model, generate embeddings, transcribe audio, or generate images without deploying anything. The free tier is rate-limited but functional for development; Dedicated Endpoints provide production-grade isolated inference with custom scaling and SLA. For agents: a single API key accesses any Hub model — prototype on the free Inference API, then graduate to Dedicated Endpoints for production. The unified REST interface abstracts across model types: text, image, audio, multimodal. Confidence is docs-derived.

REST API: POST https://api-inference.huggingface.co/models/{model_id}. Request body varies by task: text generation: {inputs: "prompt", parameters: {max_new_tokens: 200}}. embeddings: {inputs: ["text 1", "text 2"]}. classification: {inputs: "text to classify"}. image generation (SDXL): {inputs: "a photo of..."} → returns image bytes. STT (Whisper): audio bytes in body → {text: "transcription"}. Response format: JSON for most tasks; raw bytes for image/audio generation. Model discovery: GET https://huggingface.co/api/models?filter=text-generation&sort=downloads&limit=20. Dedicated Endpoints: same REST interface at custom endpoint URL.

HF API token: Authorization: Bearer {hf_token}. Tokens from huggingface.co → Settings → Access Tokens. Account-scoped. HTTPS enforced. No OAuth2. Free tier: rate-limited (requests per hour per model, varies by model size). Dedicated Endpoints: require active HF subscription; endpoint-specific access. Private models: require token with read permission for that model/org. Model gating: some models (Llama 3, Gemma, etc.) require accepting license agreement on HF website before API access.

Standard HTTP status codes with JSON error bodies. 503 "Model is currently loading" on first request (cold start for serverless models; retry after est_time seconds). 429 for rate limit (free tier). 403 for gated models not yet accepted. Model loading time: can take 20–120 seconds for large models on free tier. Dedicated Endpoints: warm always-on instances, no cold start. Free tier reliability variable — shared infrastructure. Dedicated Endpoints: SLA available. Inference API uptime at status.huggingface.co.

huggingface.co/docs/api-inference covers task reference (all supported tasks with request/response examples), authentication guide, and Dedicated Endpoints documentation. Getting started: create HF account (free), generate access token, first inference call to any model in under 5 minutes. @huggingface/inference JavaScript SDK and huggingface_hub Python library for idiomatic integration. Model cards on Hub include inference API examples. Community via HF Discord (100k+ members), Forum, and GitHub. Active ecosystem with thousands of community tutorials and integrations.