Chat Completions

The primary inference endpoint is Chat Completions, served at POST /api/v1/chat/completions. It follows the OpenAI Chat Completions wire format, so OpenAI SDKs and most OpenAI-compatible tooling work unchanged.

Request

Field	Type	Notes
model	string	A model ID from the catalog.
messages	array	Conversation so far (role + content).
stream	boolean	Set true for server-sent token streaming.
temperature	number	Sampling temperature.
max_tokens	integer	Cap on generated tokens.

Supported fields vary by model

The fields above are widely supported, but they are not guaranteed for every model, and many models accept more than this. Each model declares its own parameter set — parameters a model doesn’t declare are dropped before the request reaches it, without an error.

Check Model parameters before relying on any option beyond model and messages.

{
  "model": "zai/glm-5.2",
  "messages": [
    { "role": "system", "content": "You are a helpful assistant." },
    { "role": "user", "content": "Explain sovereign AI in one sentence." }
  ]
}

Response

A standard chat.completion object; the reply is in choices[0].message.content, with token counts under usage.

Other endpoints

Chat completions is the most used endpoint, but not the only one. The same API key and base URL also serve:

Endpoint	Purpose
POST /embeddings	Vector embeddings.
POST /rerank	Reranking, Cohere-style (query + documents).
POST /reranking	Reranking, NVIDIA NIM-style (query.text + passages).
POST /images/generations	Image generation.
POST /audio/transcriptions	Speech to text.
POST /audio/translations	Speech to English text.
POST /completions	Legacy (non-chat) text completion.
GET /models	List the models available to you.

Which rerank endpoint to use is determined by the model, not by preference — each reranker supports one of the two schemas. See the API reference for the full schema of every endpoint.

Wire formats we do not serve

We serve the OpenAI wire format. The OpenAI Responses API (/responses) and the Anthropic Messages API (/v1/messages) are not served — tools that speak those formats need a translating proxy.