Inference API: Docs

Base URL and keys

Every instance is served at:

https://chat.syndicai.dev/<your-instance>/v1

The name is shown on the instance page. Authenticate with an API key you create there:

Authorization: Bearer your-api-key

The model field is the model the instance serves, shown on the instance page (for example zai-org/GLM-OCR). GET /v1/models lists it. An instance serves one model.

Chat and vision: `POST /v1/chat/completions`

The standard OpenAI chat completions request, with streaming ("stream": true). See the API reference for the common fields.

curl https://chat.syndicai.dev/support-bot/v1/chat/completions \
  -H "Authorization: Bearer your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-oss-120b",
    "messages": [{"role": "user", "content": "Summarize this ticket in two lines."}],
    "max_tokens": 200
  }'

/v1/completions and the Anthropic-style /v1/messages are served too.

Embeddings: `POST /v1/embeddings`

On an embedding instance:

curl https://chat.syndicai.dev/search-embed/v1/embeddings \
  -H "Authorization: Bearer your-api-key" \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen/Qwen3-Embedding-0.6B", "input": ["first text", "second text"]}'

The response has one vector per input in data[i].embedding, in the order you sent them. Send several inputs per request to use the GPU well. Embedding instances answer only embeddings requests; chat models answer only chat requests.

Images and OCR

Image input uses image content parts in a chat request. Give the image as an HTTPS URL or as a base64 data URL.

curl https://chat.syndicai.dev/ocr-pages/v1/chat/completions \
  -H "Authorization: Bearer your-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-OCR",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64>"}},
        {"type": "text", "text": "Read this page and return it as markdown."}
      ]
    }]
  }'

For a URL, use {"url": "https://example.com/page.png"}. OCR models read page images, so convert a PDF to one image per page first. The model returns text, usually markdown.

Image limits


Image input	Must be on for the instance (always on for OCR models; on by default for inference instances)
Request body with image input on	Up to 64 MiB. With it off, 20 MiB and image requests are not accepted.
Images per request	Up to 4
Cost of an image	About 280 tokens on Gemma 4, whatever its resolution. Other models differ, and OCR pages cost more because they are read in detail.
Large bodies	A body over 8 MiB is passed to the model as sent, without the instance's default request settings (for example the thinking token budget). Put such settings in the request itself.

Image errors

Case	Response
More than 64 MiB	`413` with `"code": "payload_too_large"`
An image sent to an instance with image input off	`400`, `"code": "image_input_disabled"`: turn image input on in the instance settings (the instance restarts to apply it)
An image sent to a text-only model	`400`, `"code": "image_input_not_supported"`: send text only or use a vision model
More images than the limit	A client error from the model naming the limit

Errors

Errors use the OpenAI envelope: {"error": {"message", "type", "code"}}. See Inference Limits for 429 and 503, and Billing for what happens when credit runs out.

Inference API