Public AI joins Hugging Face Inference Providers
Hugging Face
- Infrastructure
- Developer tools
- #hugging face
- #inference providers
- #vllm
- #public ai
Browse Infrastructure topic page →
Hugging Face announced that Public AI is now a supported Inference Provider on the Hub, letting users call public and sovereign models through model pages or the JS and Python client SDKs. The post describes a vLLM-based, OpenAI-compatible backend routed via a global load balancer, notes free usage at time of writing, and links to documentation for setup.
What Hugging Face announced
In a September 17, 2025 post, Hugging Face introduced Public AI as a supported Inference Provider on the Hub. Users can pick it on compatible model pages or pass provider="publicai" in the Python huggingface_hub InferenceClient or the JS @huggingface/inference client; the example model is swiss-ai/Apertus-70B-Instruct-2509. Hugging Face describes Public AI as a nonprofit, open-source utility whose inference runs on vLLM servers exposing OpenAI-compatible APIs, spread across clusters donated by national and industry partners and fronted by a global load balancer. Requests can use a custom Public AI key or be routed and billed through Hugging Face. The post says usage was free at time of writing and that pricing may change; an author comment adds a 20 requests-per-minute cap on Public AI's own free keys.
Scenario: a public-sector developer pilot
Consider a developer at a municipal or research organization who needs to test a sovereign model such as Apertus-70B without committing to a hosted contract. A team could try adding Public AI as a provider option in an existing Hugging Face SDK client, keeping the provider as a configurable parameter rather than hard-coding it. The plausible benefit, per the announcement, is access to institution-built models through one familiar interface. Treat that as a starting hypothesis: the excerpt gives no latency, throughput, or reliability data, so the pilot should establish whether routing through Public AI meets the team's own service-level expectations before any broader rollout.
Trial steps and success measurement
A team could try a small evaluation: pick two or three prompts representative of the intended workload, record results and latency on the current provider as a baseline, then repeat against Public AI with the same prompts and a recent SDK version (the post cites huggingface_hub >= 0.34.6). Compare output quality, latency, and rate-limit errors, and log which mode was used, since custom-key and Hugging-Face-routed requests take different billing paths. Failure checks should include behavior at the stated 20-requests-per-minute ceiling and visibility of errors when a provider is unavailable. Because the text offers no benchmark, any latency or quality threshold is the organization's own; a human reviewer should validate outputs before relying on them in any decision-making workflow.
Trade-offs and when not to use it
The main attractions are one SDK interface and, at announcement time, free access; the caveats are equally concrete. Pricing and availability may change, and the free tier's request cap from the author comment suggests it suits exploratory or low-volume work rather than sustained production traffic. Because Public AI routes across donated clusters in multiple countries, users with data-residency or compliance requirements should confirm where requests are served before adopting it. It is also not a substitute for evaluating model quality: the post lists supported models elsewhere but the excerpt provides no comparative benchmark. If a team needs contractual guarantees or measured throughput now, this announcement alone is insufficient justification.
Context and limitations
- Based on a short source excerpt from a blog post; the full documentation and model listings were not analyzed.
- Free pricing and the 20-requests-per-minute figure come from the announcement and an author comment, which note that pricing and availability may change.
- No independent performance, reliability, or latency testing is reported in the supplied text.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.