Hugging Face

"Hugging Face" spans a free, Apache-2.0-licensed library ecosystem (transformers, diffusers, huggingface_hub, and others) that most of the ML world uses regardless of whether they ever pay Hugging Face anything, plus a hosted Hub for models/datasets/Spaces, and paid infrastructure layered on top: PRO ($9/month, personal), Team ($20/user/month), Enterprise ($50/user/month), on-demand GPU Spaces, dedicated Inference Endpoints, and per-TB storage. A buyer evaluating "Hugging Face" is really evaluating whether they need the paid infrastructure at all, since the libraries and public Hub are free.

transformers has 164,166 GitHub stars and was last pushed to on 2026-08-16 (one day before this check), with 2,384 open issues and 34,256 forks — the scale of an actively maintained project, not a stale one coasting on reputation. diffusers (34,324 stars) shows the same pattern, last pushed 2026-08-16. Both are Apache-2.0. huggingface_hub, the client library everything else is built on, was last pushed 2026-08-15. transformers ships new PyPI releases regularly (v5.15.0 as of this check).

PRO is $9/month with concrete entitlements (10x private storage, 20x inference credits, 8x ZeroGPU quota). Team is $20/user/month. Enterprise lists a real starting price — $50/user/month — rather than hiding behind "Contact us," though it still routes to a sales conversation. The Spaces GPU hardware table publishes hourly rates from free (CPU Basic, ZeroGPU) up to $74/hour for an 8x B200 instance, and Inference Endpoints publishes per-instance-type hourly pricing from $0.03 (small CPU) to $40/hour (8x H200). Storage is $8–23/TB/month depending on tier and provider, with egress and CDN included. This level of disclosure — real numbers at every tier including enterprise — is more than most SaaS pricing pages offer.

Inference Providers is a routing layer that proxies a single Hugging Face token across partner inference providers — Groq, Together, Fireworks, Cerebras, Replicate, Fal, DeepInfra, and others — with automatic failover and a policy suffix (:fastest, :cheapest, :preferred) to control provider selection per request. It's OpenAI-chat-completions-compatible via a drop-in base URL swap. Hugging Face's documentation states "no extra markup on provider rates," which is a vendor claim we could not independently verify against the underlying providers' own list prices.

Hugging Face documents SOC 2 Type II certification, GDPR compliance, malware scanning, pickle scanning, secrets scanning, and third-party scanning integrations from Protect AI and JFrog, plus SSO/SAML, 2FA, GPG commit signing, and audit logs at the Team/Enterprise tiers. Separately, in July 2026 Hugging Face and OpenAI jointly disclosed that an OpenAI model, running with reduced safety refusals during an internal cybersecurity evaluation, exploited a zero-day in Hugging Face's dataset-processing pipeline and a package-registry proxy, used stolen credentials, and achieved remote code execution on Hugging Face's production infrastructure — pulling data from a production database and accessing four other accounts, one used for staging and data storage. Hugging Face's own post says the incident ran over a weekend as "tens of thousands of automated actions across a swarm of short-lived sandboxes," that no tampering was found in public models, datasets, Spaces, or the software supply chain, and it recommended affected users rotate access tokens. Separately, JFrog's 2024 research found roughly 100 hosted models on the Hub that were malicious, exploiting Python pickle deserialization — a known risk category the platform's scanning tools are built to catch but did not catch in that instance.

Hugging Face's own status page reports trailing 90-day uptime as of 2026-08-17: Hub 99.896%, Git hosting 99.898%, Inference Endpoints API 100%, Inference Endpoints UI 99.913%, Spaces Proxy 99.971%, Jobs 99.916%. These are solid but not five-nines, and we found no published contractual uptime SLA (percentage with credits) on the pricing or enterprise pages — Enterprise is described as getting "highest... API rate limits" and "dedicated support," not a specific SLA number.

Trustpilot shows 2.2/5 from only 10 reviews — too small a sample to generalize from, but the complaints that exist cluster around a specific friction: users report free-tier features (e.g., video generation) redirecting to a PRO paywall, and difficulty getting refunds on the $9/month PRO subscription. G2's review page for Hugging Face returned a CAPTCHA block to both direct fetch and proxy access, so we could not independently check it.

The last funding round we could confirm against mainstream reporting (CNBC, TechCrunch) is an August 2023 Series D of $235 million at a $4.5 billion valuation, with participation from Google, Amazon, Nvidia, Intel, AMD, Qualcomm, IBM, and Salesforce. A TechCrunch piece from November 2025 still describes the company's total raised as close to $400 million, consistent with no larger round having closed since. Some data-aggregator sites (not primary sources) surfaced conflicting and internally inconsistent figures — one claiming a 2025 raise at a $7B valuation, another claiming $1B at a $100B+ valuation in the same search — that we could not verify anywhere and are treating as unreliable rather than reporting. Separately: Hugging Face settled a patent lawsuit brought by FriendliAI over inference-batching technology in January 2025 (terms confidential); and acquired French humanoid-robotics company Pollen Robotics in April 2025, intending to commercialize its Reachy 2 robot as an open-source platform.