Premium (private GPU)

Premium is a separate lane from Playground and serverless Models.

Use a GPU (you pay): rent a real GCP Cloud Run SKU (L4 / A100 / H100) or Use this Mac locally (Metal / Studio / Ollama) at $0. That is self-use — you are not renting from yourself.

Share your GPU (you earn): if this host is powerful enough, list it at a real $/hr so other OpenDoor orgs can run jobs while you are away. Earnings accrue in cents on gpu_host_shares and the premium_rentals row.

It does not send traffic to shared Vertex Imagen / Model Garden MaaS. Media playground (/dashboard/playground/media) is a different path.

What must be running

Image gen uses Studio: /api/studio/* → gateway POST /v1/images/generations. Optional PRIVATE_IMAGE_GEN_URL is an OpenAI-compatible worker (POST {base}/v1/images/generations). If nothing is configured, generate returns 503 { "error": "Studio GPU offline" }.

Dashboard → Studio (/dashboard/studio) is the image UI. Premium is the rental / attach / host-share flow.

Ollama on :11434 (OLLAMA_HOST) is optional for chat on this Mac and is part of share eligibility when Studio is down.

Use from the UI

1. Open Dashboard → Premium.
2. Under Use a GPU, pick Use this Mac or a GCP SKU — or attach an existing dedicated GPU.
3. Choose a catalog model (FLUX.1 Schnell is live on the private image process; MiniMax H3 is available_on_request).
4. Optional hours, or leave running until you stop.
5. Start. Local self-use says Use this Mac, not “Rent This Mac”.

Share / earn demo

1. Under Share your GPU, list this Mac if it is eligible (Apple Silicon or a discrete GPU, enough memory, Studio or Ollama up). Site admins can create a demo listing when the machine does not qualify.
2. Available hosts shows real gpu_host_shares rows. Select one and Rent listed host, or Preview as renter on your own listing.
3. The host row shows in use and earnings. Unlist only when idle. The renter (or host) can Stop.

Remote dispatch to another person’s machine is not a live marketplace yet. The listing, rental, and earnings ledger are real.

Billing

GCP GPU-seconds come from gpu_skus (same compute-billing job). Use this Mac is $0. Shared Metal uses the host’s listed hourly rate. Site-admin unlimited applies to paying renters, not to wiping host earnings.

Models

Do not call minimax-h3 as a serverless chat id. Use premium:<rentalId> or custom:<deploymentId>.