CLI hosts
You pick a model, see what it pays on your Mac, and host it from the CLI. Hosting happens in the CLI, not on the web: pulling, verifying, and serving a model is an on-device job, so this page has no "host" button. Browse below, then run the command for the model you want. Models added via the CLI then appear in Models you host. Your device must be online for it to serve and for its state to show as live.
Your hardware
…
…
1h24h
$0$0.50
What it pays · top model …
net / month · at 100% utilization · local estimate
$0
on-demand $0 / M tokthroughput …base-pay floor $0/moelectricity $0/hr
$
umbra hostThe headline is a ceiling: what this Mac earns serving requests 24/7 at 100% utilization. A single Mac rarely runs flat-out, so real earnings depend on demand. On-demand pay is per token you generate; below it, you still earn a hardware-tiered base-pay floor just for being attested-and-ready. You're paid
max(on-demand, base). Provider share is 100% in alpha, ~90% post-launch. Numbers are live from coordinator when reachable; otherwise a matching local estimate is shown (see the label above).Models that fit live catalog · most profitable highlighted
| Model | Fit | tok/s | Earn / M tok | Net / mo @ 100% | Host from CLI |
|---|
Prices are set by the platform, not by you. To host a model, copy its
umbra host command and run it in the CLI on a Mac that is signed in (umbra login) and online.Filter by model only models that fit your hardware are shown
Memory fit: a model needs
params × bytes_per_param + 1.5 GB of unified memory · budget is 80% of total. Throughput uses a bandwidth heuristic: tok/s ≈ chip_bandwidth_GBps / model_weights_GB × 0.8 (Apple-silicon decode streams the weights once per token; capped at 300).