Pick a GPU and an open-weights model. In a couple of minutes you have an OpenAI-compatible endpoint on a machine nobody else can reach, billed by the minute, that turns itself off when you stop talking to it.
One tenant per GPU. No shared inference, no batching with strangers' prompts, no queue. Your machine, your VRAM, your context window.
Each machine runs one model on a production inference server with paged attention and prefix caching. Pick from the suggestions or bring any model that fits the VRAM.
/v1/chat/completions, /v1/models, streaming. Anything that speaks the OpenAI API works: the launcher, Hermes, your editor, curl.
The meter runs from boot to shutdown, by the minute. Try a model for ten minutes and pay for ten minutes.
Set an idle timeout and a maximum runtime when you create it. No requests for that long and the machine stops itself. Restart it with one call.
Qwen/Qwen3-8BQwen/Qwen3-14BQwen/Qwen3-32BNousResearch/Hermes-4-70BQwen/Qwen3-235B-A22B-Instruct-2507-FP8| Machine | VRAM | Per hour | Per minute | Suggested model |
|---|---|---|---|---|
| Starter 16 GB | 16 GB | $0.75 | $0.0125 | Qwen/Qwen3-8B |
| Compact 24 GB | 24 GB | $1.00 | $0.0167 | Qwen/Qwen3-8B |
| Standard 24 GB | 24 GB | $1.40 | $0.0233 | Qwen/Qwen3-14B |
| Performance 48 GB | 48 GB | $2.45 | $0.0408 | Qwen/Qwen3-32B |
| Pro 80 GB | 80 GB | $3.15 | $0.0525 | Qwen/Qwen3-32B |
| Frontier 80 GB | 80 GB | $4.95 | $0.0825 | NousResearch/Hermes-4-70B |
| Frontier 141 GB | 141 GB | $5.70 | $0.0950 | NousResearch/Hermes-4-70B |
| Titan 180 GB | 180 GB | $7.80 | $0.1300 | Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 |
Pro takes 10% off, Studio 15%. Per-minute price is the hourly price divided by 60. GPU access starts with the Plus tier. Full pricing.
The launcher keeps a registry of places work can go: your local GPU, providers you hold keys for, and any OpenAI-compatible endpoint. Register your private GPU machine once and every agent, including Rix, can pick it from the model dropdown or delegate to it.
Nothing in the launcher needs to know what is behind the URL. Stop the machine and the backend simply reports unavailable until you start it again.
# create in the dashboard or with the CLI (preview)
$ omarchy-agent-launcher cloud gpu create qwen32 \
--machine performance-48 \
--model Qwen/Qwen3-32B --idle 10m --max 4h
✓ gpu_… starting · endpoint ready in ~2 min
base_url: https://api.omarchy.fans/v1/gpu/gpu_…/openai/v1
key: ofg_… (shown once)
# register it as a backend, once
$ omarchy-agent-launcher backends add qwen32 \
--kind endpoint --base-url "$BASE_URL" --key "$KEY"
# now any agent can use it
$ omarchy-agent-launcher delegate --backend qwen32 \
--task-title "Refactor the parser" --waitYour prompts are never batched with anyone else's. The model weights are loaded for you and unloaded when you stop.
The inference server's logs live on the machine and die with it. Our relay records timing and byte counts for billing, never prompt or completion text.
Each machine has its own key. Rotate it any time. The machine's own address is never exposed; the only door is our relay, and only your key opens it.
You can bring a LoRA adapter for a supported base model and we will load it next to the weights. A full tuning pipeline is on the roadmap; today, tuning happens on your side, serving happens here.
Between one and three minutes, depending on model size; the weights are cached near the GPUs. You are billed from the moment the GPU is attached, so the first minutes count.
Yes, up to a machine's count limit, for models that don't fit one card. The price multiplies by the count.
No. A stopped machine returns a clear "stopped" error instead of quietly starting a meter. Start it from the dashboard, the CLI, or the launcher.