The mother model

One larger model (6 layers, 6 heads, 192 dimensions, 4,096-token vocabulary) trained on a rolling sample of up to 4 MB from every coin's filtered corpus and the lab packs; once there are 10 or more sources each is capped at 10% of the sample. It gets 5% of every coin's fees at its own address.

Chat opens for wallets holding enough $TRAIN once its first checkpoint exists: sign a message (no transaction) for a 24 hour session, up to 10 messages per 10 minutes, prompts up to 200 characters, replies of 96 tokens, filtered. It is a small model trained here from random weights and it is often wrong.