Live run
Network
One model, cut into pipeline stages and trained by everyone at once. Watch the loss fall, see which nodes serve each stage, and check that their work verifies.
Nodes online
Tokens trained
Current loss
Epoch reward pool
trade fees · SOL
Epoch
Nodes online vs model floors
Below its floor, a model size can't keep every stage served.
1B config llama_1B_C… / 24
8B config llama_8B_C… / 96
Training loss
Cross-entropy on held-out tokens, reported by the coordinator. Orange: moving average.
Throughput
Tokens per second through the full pipeline. More nodes per stage means more parallel microbatches.
Verification
Spot checks re-run a random sample of a node's work. One failure zeroes that node's points for the epoch.
Pipeline
The model is cut into consecutive stages and each node serves exactly one, so no machine holds the full weights. Compressed activations flow forward, gradients flow back, and replicas of a stage average with each other.
onlineofflinejailedactivationsgradients
Nodes
Every registered node, its stage, verification record and points. Click a node for its history.