centram
live demo auto-loops · 60s ← Back to Home
centram in 60 seconds

how peer-to-peer
AI actually works

six scenes · auto-playing · loops
node A your laptop node B friend's machine node C phone in pocket
node A · model + LoRA
embedding
layer 0
layer 0 · lora_A · lora_B
layer 1
layer 1 · lora_A · lora_B
layer 2
layer 2 · lora_A · lora_B
layer 3
layer 3 · lora_A · lora_B
lm_head
training loss↘ converging
iter 0 iter 250 iter 500
node A · full model
842K params total
extract LoRA
256KB
genome packet
65,536
params 7.22%
of model
A sender B receiver ACCEPTED
input
prompt
"the quick brown fox "
shard A · port 9801
layers 0 — 1
first half + embedding
shard B · port 9802
layers 2 — 3
second half + lm_head
generated sequence (live)
equivalence check

verdict: pass

baseline:[106, 117, 109, 112, 115, 32, …]
distributed:[106, 117, 109, 112, 115, 32, …]
match:true (bit-exact)
try it yourself: experiments/hyperlink_mvp/
1

centram in 60 seconds

Six scenes show how local LLMs train on your hardware, exchange weight-delta genomes peer-to-peer, then run distributed inference together. The whole demo loops automatically.