Private Models

Not yet available.

Private models let you bring your own fine-tuned or self-hosted model endpoint into OpenDunes. Once registered, a private model behaves exactly like any model in the public catalog — same POST /v1/chat/completions endpoint, same DA billing, same streaming and error handling — but it is visible only to API keys in your account.

Overview#

A private model is a named entry in your account that points to an external OpenAI-compatible endpoint you control. OpenDunes proxies requests to your endpoint, applies your account's DA billing, and returns responses in the standard shape.

Use cases include:

• Fine-tuned models hosted on your own GPU server
• Models deployed in-country for Law 18-07 data residency compliance
• Internal models you want to expose to teammates via their own API keys

Registering a private model#

Registration is done through the dashboard at /dashboard/models (not yet available). You provide:

Note. Your endpoint must accept POST /v1/chat/completions in OpenAI-compatible format. Non-streaming and streaming (text/event-stream) are both supported.

Using a private model#

Once registered, call your model the same way as any other:

curl https://opendunes.com/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENDUNES_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "my-org/my-model",
    "messages": [{ "role": "user", "content": "Classify this support ticket." }]
  }'

Your model slug appears in the response model field, and requests are billed in DA at the rates you configured.

Visibility and access control#

Private models are scoped to your account. They do not appear in the public GET /api/models catalog — they are only returned when queried with a valid API key from your account. Teammates with keys under the same account can use them; external developers cannot.

Pricing (DA)#

Billing for private model requests works the same as public models: input and output tokens are charged at your configured DA rate. Your DA balance is debited per request in micro-DA.

Data flow and residency#

Requests to private models are proxied through OpenDunes' infrastructure before reaching your endpoint. If your endpoint is hosted in Algeria, the full request path stays within the country. If you need strict end-to-end residency for Law 18-07 compliance, ensure your endpoint's host is within an approved jurisdiction and that your OpenDunes account has data residency mode enabled (not yet available; ships alongside this feature).

Limitations#

• Private model endpoints must be reachable from OpenDunes' outbound IP ranges (published in the dashboard).
• Maximum response timeout is 120 seconds. Long-running generations should use streaming.
• Private models cannot be used with Auto Exacto model selection — accuracy scores are only available for models in the public catalog.