The measured one
Private
A raw hop still reconstructs the prompt. One MPC share is noise. The bound costs 2.42 MiB per token. The figure is in the Paper.
Meet me behind
the old library
at midnight.
recover 100%
0.21% · chance
Private inference network
Decentralized inference on contributed GPUs. Our research goal: a network where an untrusted node cannot recover what you typed.
Share lattice · Game of Life
01 / The problem
Split a model across strangers' GPUs and every node still touches raw activations - enough to reconstruct pieces of a prompt.
02 / How it leaks
Every node handles real activation tensors. SILQ exists to quantify that exposure - and close it.
The measured one
A raw hop still reconstructs the prompt. One MPC share is noise. The bound costs 2.42 MiB per token. The figure is in the Paper.
Meet me behind
the old library
at midnight.
recover 100%
0.21% · chance
Anyone's GPU, pooled into swarms. No datacenter.
Any model. No content-policy layer in the way.
Privacy isn't a feature to bolt on later. It's the open problem - and the one SILQ exists to close.
03 / Public vs private
private run
one share is noise
0.21% · chance
04 / Product
Background GPU via Ollama. Larger models, higher pay per job.
Paid per job
npx -y @silqxyz/worker@0.2.0 --token YOUR_TOKENGet install commandIn-tab inference via WebGPU. One click, no terminal.
Paid per job (lower)
Open browser workerDrop-in chat completions. We don't store prompts; 1 credit per request.
1 credit / request
curl https://api.silq.example/v1/chat/completionsTry the playgroundPrivate inference economy
What crosses the chain is payment. Your prompts never go on-chain.
We keep no prompt history or profiles. In v1 the serving worker sees the prompt.
The margin on every paid job funds the network, its privacy research and $SILQ buybacks.
SILQ