Every model here is an open weights build served by peers on their own hardware. Providers advertise a price range and settle the exact rate over an encrypted channel, so what you see is the current network median.
If your machine can hold the weights, the app runs the model on your own GPU. No payment, no peers, no network involvement at all.
Route to a provider with the hardware for a 70b. Encrypted to their key, computed in a sealed VM, returned with a proof.
Pull a model and serve it. If peers request it, you get paid for it at whatever rate you advertise.
Providers publish a range rather than an exact number, which keeps demand from being read off the network. Your client negotiates the final rate inside the encrypted channel, then locks that amount in escrow before any work starts.