Run AI models at lower costs.

One OpenAI-compatible endpoint for open-weight models running on idle consumer GPUs. Swap your base URL, keep your code — once the network is live.

One API, familiar shape

POST http://localhost:8787/v1/chat/completions
Authorization: Bearer nf_demo_<your-key>
Content-Type: application/json

{
  "model": "qwen3-8b",
  "messages": [{ "role": "user", "content": "Hello" }],
  "stream": false
}

Chat completions, text generation, streaming where supported, token usage reporting, model selection, API-key auth and rate limiting (demo: 60 requests/minute per key). Full examples in the documentation.

Open-weight models

Qwen3, Llama 3.1 and Mistral launch candidates — small quantised models that fit consumer GPUs. Licences checked before serving; see models.

Usage reporting

Every response reports input and output tokens, so you can reconcile cost against the published per-million-token prices.

Playground first

Try a model in the browser before writing code — the chat playground on the models page talks to the same demo endpoint.

Who it is for

  • Independent developers and AI agent builders who need cheap inference for prototypes and side projects.
  • AI startups and SaaS companies that want a second, lower-cost inference option.
  • Researchers and open-source projects working within tight budgets.
Pre-launch, honestly: the API is in demo mode on localhost. There are 0 live workers and no models serving live traffic yet. Early access means testing the demo, shaping the API, and being first when the beta opens — not production traffic today.

ACCESS AI COMPUTE Read the docs