Docs
How models are hosted
You always call one endpoint with one key, in the OpenAI format. Behind each model id sits a route that says where the model actually runs. This page explains those routes, so you know what happens to a request after it reaches the Club.
One id, one route
Every public model id (like deepseek/deepseek-v4-flash) maps to exactly one route. A route records:
- Kind: chat, embedding, image, audio or rerank. The kind decides which endpoint accepts the model.
- Provider type: which adapter talks to the upstream server (see below).
- Base URL and upstream model name: where the request goes, and what the model is called there.
- Upstream key: stored as a server environment variable. The route only names the variable; the secret never lives in the database or reaches your browser.
- Catalog facts: list price, closed or open-weight, Guest availability and the Member and Executive prices for Private and Research Share.
Club admins add and edit routes in the admin area at /admin/models. Changes apply within seconds, without a deploy.
Model kinds and endpoints
| Kind | Endpoint |
|---|---|
| Chat | POST /v1/chat/completions (streaming supported) |
| Embedding | POST /v1/embeddings |
| Image | POST /v1/images/generations |
| Audio, rerank | Listed in GET /v1/models. No public endpoint yet. |
Calling a model on the wrong endpoint returns 400 with code: unsupported_endpoint.
Provider types
| Provider type | What it is | Supports |
|---|---|---|
openai-compatible | Any hosted API that speaks the OpenAI /v1 format. Requests pass through as-is. | Chat, embeddings, images |
vllm, ollama, tgi | Self-hosted inference servers for open-weight models, using their OpenAI-compatible /v1 APIs. Same adapter as above, with each server's usual local base URL as the default. | Whatever the server exposes (usually chat, often embeddings) |
gemini | Google's OpenAI-compatible endpoint for Gemini models. | Chat, embeddings, images (where the upstream offers them) |
anthropic | Native adapter for the Messages API. It translates OpenAI chat requests (system prompts, images, tools) to the native format and translates responses and streams back. | Chat only |
What stays the same for you
- Same format everywhere. Responses and streams come back in the OpenAI format whatever the upstream speaks.
- Same billing. Usage is metered from the token counts the upstream reports (estimated when it doesn't report them), and priced as a percent of that model's list price for your tier and key. See Pricing.
- Same data policy. Your key's policy applies at the Club, before anything reaches the upstream. Upstream providers may keep requests briefly for abuse monitoring under their own terms, unless the model shows a ZDR badge in the catalog.
- Same errors. Upstream problems map to the error types in the API reference, without leaking upstream details.
Regions
Each model is served only in the regions its provider supports. The catalog shows regions per model (TBD until confirmed).