qiubit

The model runs on the blockchain. The stories260K weights (260K parameters) live in an Octra program, and the node computes every token in Q16 fixed point. Outputs match the Python reference bit for bit.

Program
code_hash
Limit: 10 positions per message, shared by the prompt and the reply. No memory between messages.
License: stories260K & llama2.c (MIT, Andrej Karpathy), trained on TinyStories (CDLA-Sharing-1.0)

How should the story begin?

Write the first words. An Octra node writes what comes next.

Story opening

3

qiubit runs stories260K on an Octra devnet node. Prompts and replies are public.