About Cria
Who runs this, what it is, and how to reach a human.
Who
Cria is built and operated by Mike, a solo founder. It is not a token, not a fund, and not for sale. It is in invite-only beta as of 2026-09.
What
An open-source (MIT) marketplace for open-model inference on independently operated hardware. People who already run Ollama, LM Studio or vLLM connect their machine with one outbound command and set a price per 1M tokens. Developers get an OpenAI-compatible endpoint and pay per token, no minimums. Cria routes, bills, and pays out, and keeps a fee from the provider's side.
The reason it exists: the big inference clouds host a few hundred popular models on datacenter GPUs. Everything else — fine-tunes, niche quantisations, 70B models on Mac Studios, the model you trained last week — has nowhere to be served as an API. The people who already run those models have hardware sitting idle. Cria connects the two without asking anyone to open ports, run containers, or hold a token.
What it is not
- Not a GPU rental. Your machine stays yours; you serve tokens from models you chose.
- Not private. Requests run on other people's computers; providers commit to confidentiality in the terms, but treat it like any third-party API, not like your own box.
- Not an SLA. Independent servers come and go; routing works around it, nothing guarantees it.
Contact
- Email: [email protected]
- Source and issues: https://github.com/EpilogueLabs/Cria
Abuse reports and takedowns are handled first. Security issues: email with "security" in the subject and you will get a reply from a person.