Public AI Inference Utility

license

MIT(developer portal and API gateway code)

source

partial(the published repository is the developer portal and gateway, which forwards requests to an internal endpoint

access

hosted OpenAI-compatible API(api.publicai.co)

engine

vLLM(on partner-donated clusters)

Public AI publishes the front door: an MIT-licensed developer portal and gateway that handles sign-up, API keys, billing and rate limits, then passes each request to an internal endpoint. The model servers and the routing across partner clusters behind that endpoint are not in the repository, so the published code cannot stand the service up on its own.

• https://cdn.jsdelivr.net/gh/forpublicai/platform.publicai.co/ recorded 2026-10-13

  The repository tree at its head holds config, docs and modules folders plus LICENSE, package.json, README.md and zuplo.jsonc; the modules are gateway handlers for API keys, billing, wallets, pricing, rate limits and request headers, and no model-serving code.

• https://cdn.jsdelivr.net/gh/forpublicai/platform.publicai.co/config/routes.oas.json recorded 2026-10-13

  The gateway's route file forwards requests with Zuplo's urlForwardHandler to the base URL "https://api-internal.publicai.co" and rewrites chat and completion calls to that internal host, whose implementation is not in the repository.

• https://cdn.jsdelivr.net/gh/forpublicai/platform.publicai.co/LICENSE recorded 2026-10-13

  The repository LICENSE is the MIT text: "MIT License" and "Copyright (c) 2025 Public AI Inference Utility", with no added terms.

• https://cdn.jsdelivr.net/gh/forpublicai/platform.publicai.co/README.md recorded 2026-10-13

  The README calls the repository "The developer portal and API gateway for the Public AI Inference Utility", for creating API keys, viewing documentation and monitoring usage and billing, built on Zuplo and deployed with "A Zuplo account for deployment"; its License section names MIT.

• https://publicai.co/about recorded 2026-10-13

  The about page says "The Public AI Inference Utility is a nonprofit, open-source project" fiscally sponsored by Metagov.

• https://publicai.co/stories/utility recorded 2026-10-13

  The launch post says inference "is handled by servers running OpenAI-compatible APIs on vLLM", deployed across clusters donated by national and industry partners, behind "A global load-balancing layer".

Public AI publishes no request, token or user count. What it does report is standing: it is a listed inference provider on Hugging Face, the official international deployer of the Swiss Apertus models, and the host for national models such as Spain's ALIA and EuroLLM, which places it as real but specialist infrastructure for public-sector models.

• https://huggingface.co/docs/inference-providers/providers/publicai recorded 2026-10-13

  Hugging Face's Inference Providers documentation has a PublicAI page for chat completion, reading "All models supported by the PublicAI Inference Utility can be found here", with an example calling speakleash/Bielik-11B-v3.0-Instruct through the Hugging Face router.

• https://publicai.co/stories/amazon-intel recorded 2026-10-13

  A June 16, 2026 post says that after launch "Traffic ramped faster than expected" and that Intel-backed clusters have hosted EuroLLM, "Spain's ALIA models" and Apertus 8B; it gives a concurrency target of about 2,500 users but no measured usage figure.

• https://publicai.co/stories/apertus-1-5 recorded 2026-10-13

  A July 24, 2026 post says "Public AI is proud to be the official international deployer for Apertus" and that Apertus 1.5 is live on its chat app, developer platform and Hugging Face.

• https://publicai.co/stories/huggingface recorded 2026-10-13

  A September 16, 2025 post announces that the Utility "is now an official inference provider on Hugging Face".

The service runs its models on vLLM, so its serving techniques are that engine's, and what it adds is routing, failover and per-key rate limits across clusters donated by different partners. Its catalog is a handful of national models of up to 70 billion parameters, and it publishes no throughput or latency figures.

• https://platform.publicai.co/inference-partners-faq recorded 2026-10-13

  The partners FAQ describes "a unified API that routes requests to open-source models hosted by inference partners worldwide", with the gateway handling "load balancing, traffic routing, and failover"; partners contribute GPU compute.

• https://platform.publicai.co/models recorded 2026-10-13

  The Available Models page lists nine catalog entries, among them swiss-ai/apertus-v1.5-8b and apertus-v1.5-70b at 262K context, aisingapore/Gemma-SEA-LION-v4-27B-IT and Qwen-SEA-LION-v4-32B-IT at 128K, and speakleash/Bielik-11B-v3.0-Instruct at 32K, served through "a unified, OpenAI-compatible API".

• https://platform.publicai.co/plans recorded 2026-10-13

  The plans page lists four tiers with per-key limits: Free at 100 requests/minute, Plus at 200, Pro at 300 and Enterprise at 10,000.

• https://publicai.co/stories/utility recorded 2026-10-13

  The launch post says inference runs on "servers running OpenAI-compatible APIs on vLLM" across partner clusters, with "A global load-balancing layer" routing requests between them.