Offer a machine
On a machine’s card in Compute, switch on Offer to the community pool. It lends the same share you set for it — power, memory per backend, hours — and only model stages: research shards from other organizations never run on it. Pause it, change its share or switch the offer off at any time; it applies at the next heartbeat.
Use the pool
Create a key that reaches the community pool (Compute → Keys → Reaches: Community pool). With it, hetero run --across all places a model across the machines on offer right now. The pool lists machines by their node names and what they lend; their owners’ names and organizations stay private.
What each side sees
| Who | Sees | Does not see |
|---|---|---|
| The machine’s owner | That the machine served community jobs, when and for how long | Who submitted them, their files or prompts |
| The pool’s user | The machines on offer: node name, platform, what they lend | Owners, organizations, addresses before a stage is placed |
| Hostwatch | Jobs, their small files, settings and heartbeats | Hidden states and weights, which flow directly between machines |
Before you offer a machine
- A stage computes its layers on the hidden states of other people’s prompts, and holds the weights of the layers it serves. Both pass through the machine. Offer machines you are comfortable sharing that way.
- Stages accept connections only from the coordinator that placed them, and use only the memory you lend; the worker refuses layers that would not fit before loading them.
- A coordinator reaches the stages directly, so the machine must be reachable for it: an open port (
--port-base) or a shared VPN.
Before you use the pool
- Your prompts pass through other people’s machines as hidden states, and your model’s layers are fetched by them. Use your own network for private data and models.
- Machines come and go. A placement holds its machines for the run; a machine that leaves ends it, and the next run plans around the machines on offer then.