Community inference
Point your code at an OpenAI-compatible endpoint. Community machines already sitting idle run the job. Qualify yours and you get paid for the ones it finishes.
Braise is in closed beta. Providers are rewarded in Compute Credits, the beta instrument. Money payouts open to every qualified provider when the network opens.
How it works
Catalog aliases, credits, invite codes: ask and we walk through them.
An OpenAI-compatible endpoint. Same calls, different base URL. No GPU, no runtime, no local model on your side.
Once, with Ollama, LM Studio, or llama.cpp, and the model it already serves. Sharing stays off until you turn it on. It stops the moment you turn it off.
Your machine is paid for every job it finishes. Compute Credits during the closed beta. Money payouts, open to every qualified provider, when the network opens.
The hardware is bought, powered, and idle. That is what the price is built on.
Instead of
Same OpenAI-compatible base URL. The price is built on idle community machines, and you are not locked to one vendor.
You skip buying a GPU, managing models, and hitting a VRAM ceiling on your own box.
You pay for finished work. You do not keep a reserved instance warm between requests.
No listing, no price to set, no reserved slot to keep online. Qualify once and get paid per finished job.
Waitlist
Providers first. Tell us what you run and what you already spend on inference. That is how we score invitations.
FAQ
Compute Credits during the closed beta: network utility units, not money. When the network opens, credits are retired and every qualified provider is paid in money. No invitation list. No permanent two-instrument system.
We will publish a dated price index against named commercial APIs once it is measured on our own settled jobs. Until then we can only describe the mechanism: you pay for hardware that is already bought and already powered.
We will not quote a figure until we have measured a distribution of rewards per online hour, per hardware tier, over a stated window. Promising a number before that would be inventing one.
Yes. Public community providers can read the prompts they execute. Encryption protects transport, not the worker. If a network tells you otherwise, they are overselling.
The request is refused with a capacity error and nothing is billed. There is no fallback to a commercial API. Availability is whatever is actually online for that model.
A GPU or high-RAM Mac with real idle hours, and a runtime you already use. We qualify it before routing anything to it. Companies with idle GPU capacity can join too.