Hugging Face Inference Providers vs LiteLLM

Hugging Face Inference Providers 25

Hosted lines. status.huggingface.co is a Better Stack page with 90-day component history for the Hub, Inference Endpoints, Spaces, the CDNs and Jobs, but no component for Inference Providers or the router (10 of 20). With no component, there is no readable incident history for the surface agents call, and the feed's 23 items from 10 June to 7 October 2026 don't name it (5). No rate limits for the router were found; the Hub rate-limits page gives numbers for Hub APIs, resolvers and pages only (0). No 429 or overload guidance for the router was found (0). No SLA found (0). Chat completions on the router is generally available; the Responses API is labelled beta (10). Routing between providers with :fastest or :cheapest softens a single provider's outage, which this checklist doesn't score.

LiteLLM 70

Read with the local-software lines, since the gateway runs on the operator's hardware and BerriAI hosts nothing an agent calls. litellm on PyPI states Python 3.10 to 3.14, and the gateway ships as a Docker image, a Helm chart and Terraform modules, with GHCR images signed by cosign (20). 51 GitHub Actions workflows plus a CircleCI config cover unit, proxy-database, end-to-end, chaos, mutation and UI tests. The GitHub API was rate-limited for our reader on 6 November and Actions pages refused it on 6 October, so we couldn't confirm main is passing (15 of 25). The issue tracker showed 1,674 open issues on 6 November, with eight bug reports among the 15 newest, seven of them with a reply, against a project that tags several releases a day (10 of 25, down from an unchecked 12). Release notes carry breaking-change bullets, but breaking changes land in weekly minor releases on the 1.x line, so version numbers don't signal them (10 of 15). 1.108.1, past 1.0 (15).

Hugging Face Inference Providers 20

No x402, MPP or L402 (0). Per-token prices for every model and provider published without a login in the router's /v1/models and on model pages (20). No free tier since 7 October 2026, when free users lost their included monthly credits; PRO's $2.00 of credits needs a subscription (0). Sign-up, token creation and credit purchase are browser steps (0).

LiteLLM 60

Read with the self-hosted rule, since an agent uses the free gateway and Enterprise is a sales-led licence for the same software. No x402, MPP or L402 (0). MIT with no account and no card, so 20, 20 and 20 on the last three lines. Model calls cost what the operator's provider keys cost.

Hugging Face Inference Providers 62

No OpenAPI document for the router (openapi.json paths answer 404). The task reference pages are generated from typed JSON Schema specs in huggingface.js, and the router's /v1/models gives a machine-readable catalogue, so 12 of 25. llms.txt for the Inference Providers docs and a Markdown twin of every page via .md (10). Guides say what each task and policy is for, with a quick setup page for coding agents (14 of 20). Parameter tables with types for chat completion, though response_format and tool_choice show as unknown types (9 of 15). Examples in Python, JavaScript and curl on every task page, but no documented error responses for the router (9 of 15). No changelog of its own; the Hub changelog doesn't cover it and dated changes are only in the hub-docs history (8 of 15).

LiteLLM 91

Every gateway serves Swagger at /, ReDoc at /redoc and the raw schema at /openapi.json, and the docs host a public Swagger explorer (25). llms.txt, llms-full.txt and a Markdown twin of every page via .md (10). Pages say what a feature is for and where its limits are, and the key-auth page explains how permissions are inherited and where they aren't, but the reference is very large for a model to navigate (16 of 20). OpenAI-shaped request schemas from Pydantic models, with metadata, litellm_params and config blocks left as free-form objects (10 of 15). A full error reference with the envelope, every status, response headers and how to tell gateway limits from provider limits, plus examples on every page (15). Release notes per version with dates, newest v1.108.0 and v1.106.7 on 4 November 2026 (15).

Hugging Face Inference Providers 58

Router reading of the model checklist (tool use, structured output, caching, context, batch, SDKs, errors), as for openrouter. Tool calling through the OpenAI-compatible API with tool_choice, support flagged per provider in /v1/models (15 of 20). Structured output documented, support flagged per provider and false on some (10 of 15). No prompt caching guidance found, though the Enterprise usage API counts cached input tokens (3 of 15). Routes to models with up to 1,048,576 tokens of context per the model list, a judgement call since a router has no default model (15). No batch API (0 of 10). Official clients in Python and JavaScript, plus any OpenAI SDK (10). Errors not documented (5 of 15).

LiteLLM 79

Read with the API lines for the gateway an agent calls. Responses size through the usual max_tokens and streaming, and the MCP gateway can replace a large tool catalogue with four virtual tools (tool search) or filter tools by semantic relevance, which cuts context cost (20 of 25). Key, team and spend listings page and filter, though we didn't check every list call (15 of 20). Errors keep the OpenAI envelope with a string code, gateway 429s carry retry-after, rate_limit_type and reset_at, management routes answer RFC 9457 problem+json, and a table says which errors are safe to retry (20). Retries, cooldowns and fallbacks run inside the gateway and the docs say which errors to retry, but there are no idempotency keys for the management writes (12 of 20). Any OpenAI or Anthropic SDK works by changing the base URL, but the only first-party client is the Python SDK, and keys and spend need Postgres and a master key before the first call (12 of 15).

Hugging Face Inference Providers 71

Closed service under Terms of Service that name Inference Providers, with an Apache-2.0 Python client and an MIT JavaScript client (20). The security page and the docs agree on no training, no stored bodies and 30-day logs, but the privacy policy dates from 28 March 2023 and each provider's own data policy governs the rest (18 of 30). No deprecation policy found. Provider support changes and the 7 October removal of free credits are dated only in the docs repository history (5 of 20). The privacy policy lists 11 subprocessors with countries per the Inference Endpoints check, and the docs name every provider a request can reach, though not where each runs (15 of 20).

LiteLLM 63

MIT for everything outside enterprise/, which has its own licence (25 of 30). The data security page (updated 6 November 2026) and security.md agree that a self-hosted gateway sends no data or telemetry to BerriAI, and the gateway stores prompts and spend in the operator's own Postgres. We found no terms of service or privacy policy on litellm.ai for its site, trial or Enterprise licence (20 of 30). A version support policy since 29 June 2026 (four stable minor lines), deprecated pages labelled in the docs, and release notes with breaking-change bullets (15 of 20). Usage telemetry, new in v1.107.0, is off by default, lists every field it can send and goes only to a table or endpoint the operator sets. A main-branch commit of 23 October 2026 made the Admin UI chat post a hashed user id to chat-telemetry.litellm.ai, and it was removed on 24 October before any tagged release contained it, so it isn't deducted. The start-up fetch of the price map from GitHub is documented with an opt-out (17 of 20).

Hugging Face Inference Providers 79

Model reading. Fine-grained user access tokens that can be limited to the Make calls to Inference Providers permission, rotated and revoked, plus a public revocation endpoint for leaked tokens, and organisation approval of fine-grained tokens on paid plans (30). Hugging Face says it doesn't store user data for training, but data handling past the router follows each provider's policy, with no per-request filter to exclude providers that train or retain (12 of 20). Request bodies and responses aren't stored when routed and debug logs are kept up to 30 days without user data or tokens, with provider retention outside that statement (10 of 15). Usage by model and provider in account settings, an Enterprise usage API, and Team and Enterprise admins can disable providers and set a spending limit (13 of 15). SOC 2 Type 2 for the Hub, of which the docs say Inference Providers is a feature, and security.txt valid to 2030 per the Inference Endpoints check; no bug bounty found (14 of 20).

LiteLLM 72

Read with the tool checklist for self-hosted software. Virtual keys are hashed, revocable and blockable, with model, route, MCP server, budget, rate and expiry limits, and a proxy_admin_viewer read-only role exists. Rotation with a grace period is Enterprise only, so 25, less 10 because the docs pass a key as ?key= on /key/info and on the Claude Code marketplace manifest (15 of 30). allowed_routes, MCP permissions by key and team, an approval queue for submitted MCP servers and stdio MCP servers off by default from 1.104.0 (15 of 20). Guardrails can run before a model call and on MCP tool results, and an OWASP LLM Top 10 page maps each risk to a control, but prompt-injection detectors such as Lakera need an Enterprise licence and all of it is off by default (12 of 15). Request and spend logs per key in the open-source build; audit logs of admin actions are Enterprise only (12 of 15). security.md with GitHub private reporting, a bug bounty from $500 to $3,000 for P0 and P1 reports, a security announcements list, SOC 2 Type II per the data security page, CodeQL, OSV and Scorecard workflows, and advisories published on GitHub. No security.txt, and reports without a video are closed unread (18 of 20).