Skip to main content
Hivenet Inference API lets you serve supported open-weight models through managed, dedicated API endpoints without setting up and operating the inference stack yourself. You choose the location, model, serving variant, replica count, and endpoint name in the Compute console. Hivenet provisions the serving infrastructure and gives you an endpoint you can call with OpenAI-compatible clients such as the OpenAI SDK and curl.
Hivenet Inference API is part of the Compute console. Endpoints and API keys belong to the active organization and use the same prepaid organization credit balance as Compute resources.

​
How it works

The deployment flow has five main choices:
  1. Location: Choose where your Inference API models run. In the current product, the selected location applies across your Inference API models; per-model location selection is not available yet. Choosing Global lets Hivenet assign an available location, and the endpoint details show where it was placed.
  2. Model: Choose a supported model from the catalog. The available catalog depends on the selected location.
  3. Variant: Choose a preconfigured serving setup for that model. A variant can define the GPU configuration, quantization, context window, performance characteristics, and price.
  4. Replicas: Choose the fixed number of serving replicas to run. More replicas provide more request capacity and increase hourly cost.
  5. Endpoint name: Keep the suggested name or enter a label you can recognize in the console.
The console lists these resources under Deployments. Each deployment creates one callable endpoint.
Autoscaling is not available yet. You choose a fixed replica count and can change it later when capacity is available.

​
Inference API, Compute, and Hivenet Router

These products cover different parts of the inference stack. If you need an endpoint but do not want to operate vLLM or another serving runtime yourself, Hivenet Inference API is the managed option.

​
Endpoint lifecycle

The Inference API page shows the current state of each endpoint. The console says starting a model takes about 90 seconds. Charges begin once it is running.
Stop is reversible. Terminate permanently shuts down the endpoint and makes its URL unusable. After termination, Clear only removes the terminated entry from the Deployments list.

​
Endpoint details

Open Details to see the endpoint URL, model and serving configuration, replica count, region, cost information, available usage metrics, and quick-start snippets for Python, JavaScript, and curl. The endpoint URL already includes the API version path, such as /v1. The generated snippet also contains the exact model value expected by that endpoint. See Call your endpoint for request examples and Manage endpoints for lifecycle controls.

​
API keys and billing

Inference API keys are scoped to the organization rather than to one endpoint. The full secret is shown once when you create a key, so save it before closing the dialog. Disabling or revoking a key stops requests made with that key but does not stop running endpoints. Inference API is billed by the fixed model-serving capacity you run. Before deployment, the console shows estimated hourly, 24-hour, and weekly cost. Stopping an endpoint pauses serving charges; terminating it ends them once termination is complete.

​
Location and data handling

The location selector can mark a location Sold out when capacity is unavailable. It also shows the data-handling notice that applies to the selected location. Read that notice before deploying. It is the source of truth in the console for where prompts and responses are processed or stored and for the other data-handling statements associated with that location.

​
Next steps