Docs

Separate from serverless. Provision a GPU, optionally attach LoRAs, and split traffic with routers.

Dedicated deployments (Wave 3)

Separate from serverless. Provision a GPU, optionally attach LoRAs, and split traffic with routers.

Call a deployment

bash
curl "$GATEWAY/v1/chat/completions" \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "custom:<deployment-uuid>",
    "messages": [{"role":"user","content":"hello"}]
  }'

Custom weights & precision

On create (POST /api/deployments) or Request GPU form:

Field	Meaning
`weightsUri`	Hugging Face repo (overrides catalog)
`precision`	`fp16` / `bf16` / `fp8` / `int4` → vLLM `--dtype`
`minReplicas` / `maxReplicas`	Cloud Run autoscaling
`scaleToZero`	Min instances 0 when idle
`reserved`	Keep warm (`minReplicas ≥ 1`, no scale-to-zero)

Multi-LoRA

bash
# Dashboard: Deployments → open deployment → Load LoRA
# Or API:
POST /api/deployments/:id/loras
{ "name": "support-v1", "adapterUri": "org/adapter-repo" }

# Then:
model: "custom:<deployment-uuid>/support-v1"

Requires a vLLM endpoint (GCP). Local Ollama returns unsupported.

Unload: DELETE /api/deployments/:id/loras/:loraId

A/B routers

bash
POST /api/deployment-routers
{
  "name": "Pricing experiment",
  "slug": "pricing-ab",
  "targets": [
    { "deploymentId": "...", "weight": 80 },
    { "deploymentId": "...", "weight": 20 }
  ]
}

# Gateway:
model: "router:pricing-ab"

UI: Deployments → Routers.

Service tiers

Premium rentals