Docs

Dedicated endpoints.

A dedicated endpoint is a model running on GPUs reserved for your project alone. You rent the hardware by the hour instead of paying per token — worth it for steady traffic, for a fine-tuned model of your own, or when you want capacity nobody else is sharing.

Shared pool or dedicated?

Shared pool

Billed per token. Nothing to start, nothing to shut down, no idle cost. The right default for bursty traffic, prototypes, and anything you are still sizing.

Dedicated endpoint

Billed per GPU-hour for as long as the replica is up, whatever traffic it serves. The only way to run your own fine-tuned weights, and predictable at steady load.

Deploying one

1Pick the model

Any model in the Foundation catalog can be deployed dedicated, and so can any model your own training jobs produced. From the model's page, choose Deploy.

2Confirm the size

The form prefills the GPUs a replica of that model needs — that number comes from the catalog, not from you, because it is a property of the weights. A model that needs eight GPUs cannot be squeezed onto fewer.

3Wait for placement

The endpoint is queued until a node with enough free GPUs in one block is available. If it cannot be placed yet, the endpoint says so in plain terms — for example that it needs eight free GPUs while the largest free block is three — rather than sitting on a spinner. It stays queued and is placed as capacity frees up.

4Call it

Once it reads Running, the endpoint answers on the same base URL as everything else. There is no per-endpoint hostname and no path segment: the model id in the body is what selects it.

Calling a dedicated endpoint

curl

curl https://inference.legionedge.ai/v1/chat/completions \
  -H "Authorization: Bearer $LEGIONEDGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "<your-endpoint-model-id>", "messages": [{"role": "user", "content": "Hello!"}]}'

The endpoint detail page prints this command with your model id already filled in. A fine-tuned model records the weights it was built from, so the id you pass is the one the console shows on the endpoint — not the name you gave the endpoint.

What it costs

GPU-hours, at the rate for the hardware the replica lands on, times the number of GPUs it needs. The per-GPU-hour rates are on the models & pricing page.

Model	GPUs per replica
`qwen3-0.6b`	1
`qwen3-4b-instruct`	1
`kimi-k3`	8

Billing runs while the endpoint exists and is placed — pausing or deleting it is what stops the meter, not the absence of traffic.