Hero Run · Machine Yield
Turn an idle machine into $HERO.
Machine Yield pays your computer to serve the pieces of AI models that are too big to run anywhere else. Stake $HERO, run one command, and your idle hardware earns for every byte it delivers. Better hardware holds bigger models and earns more.
why it's even possible
how you earn
Every epoch, a pool of $HERO is split across the machines serving the network. Your share is what your node proves, weighted by your stake:
your share = pool × ( √stake × committed_GB × health × uptime ) ⁄ everyone's total- √stake: staking more helps, but at its square root, so a great machine with a small stake still out-earns a big stake on a weak one. Hardware stays the point.
- committed GB: how much of a model you hold hot. Bigger machines hold bigger models and earn more.
- health: the fraction of correctness checks you pass (more on that below).
- uptime: how much of the day you were actually serving.
Yield accrues to your key and you claim it as credits that spend exactly like a top-up: on inference, or anything else on Hero Run.
Proof, not promise
There are no committees deciding whether you did the work. The network asks your node for random pieces of the model and checks them byte for byte against what they should be. Serve the right data fast, or earn nothing. Correctness is deterministic, so there's nothing to fake and nothing to argue about. A machine that can't hold or serve its slice simply doesn't score.
Is my stake safe?
Your principal is returnable one-to-one, any time. Staking comes out of your key's spendable balance and goes straight back when you unstake. A slow day never costs you principal; a machine that underperforms simply earns less that epoch. Stake weighs your rewards, it isn't a bet you can lose.
start in three steps
curl -fsSL https://herorunai.com/join.sh | sh -s -- hr_live_your_keyApple Silicon runs the full expert engine. Intel Macs, Windows and Linux join as sidecar nodes: they hold and prove model data rather than serving experts, because the streaming engine is Apple-only.common questions
What hardware do I need?
Apple Silicon (M1 and later) runs the full engine: it streams experts from your SSD and holds a slice of a model in unified memory. Anything with 8 GB can serve a small model or a slice of a bigger one, and a 48–64 GB machine can hold a 100-billion-parameter model hot. Intel Macs cannot run the engine at all. There is no MLX build for them, but they can still join as sidecar nodes, as can Windows and Linux boxes. No spare machine? Rent one through Hero Metal and it auto-joins on boot.
Can my machine earn if it can't run the engine?
Yes. That is the seeding tier. Intel Macs, Windows and Linux boxes have the two things distribution needs: disk and bandwidth. They hold a model's blocks and serve them to machines that are joining, and the receiver hash-checks every block on arrival. You are paid for verified bytes delivered, reported by the machine that received them, so the work counts only when it was real.
Do I need a static IP or port forwarding?
No. The join script opens a Cloudflare quick tunnel and registers that public address for you. A node the coordinator cannot reach still earns the unreachable floor, but a reachable one can be challenged and proven, which scores far higher.
What do I actually earn?
A share of each epoch's pool, proportional to what your node proves and your stake's square root. Bigger, faster, more-reliable machines earn more. The pool size is set by Hero Run and grows as network usage does.
What if I go offline?
Uptime is part of your score, so downtime lowers what you earn that epoch, but nothing is lost. Come back online and you score again. Your stake is untouched.
Can I lose money?
No principal is at risk from serving. Stake returns one-to-one, and a bad epoch just earns less. The only cost is the electricity and bandwidth your machine already runs on.
Pick a model. Run a node. Get paid.
Start earning