PlebGPT

A tiny transformer running locally in your browser, with weights stored in a single Bitcoin transaction.

389 KB model artifact
Prompt
Advanced
Defaults match the current browser sampling settings.
slop stream
Running Client-side inference
Model not loaded yet...
What is this?
A small on-demand slop machine This is a very small transformer that generates pleb-slop on demand. Give it a prompt, and it produces compact Bitcoin-maxi claim fragments directly in the page.
Trained on low-grade signal The model was trained on a large corpus of carefully prepared plebslop: monetary grievance, proof-of-work rhetoric, sovereignty larping, and the lowest-quality Bitcoin discourse we could distill into usable training text.
Designed for a standard transaction The model was engineered around a hard size constraint: the compressed weights fit inside a single standard Bitcoin transaction, under the 400 KB standard policy limit. The current model artifact is about 389 KB.
On-chain UI, local inference The UI and the weights are both on-chain in inscriptions. When loaded from an ord content URL, the UI finds the latest child inscription containing the weights, downloads that file, decompresses it, reconstructs the tensors, and runs all inference locally in JavaScript. There is no hosted model API in the generation path.
Model Technical Details
Vocabulary Size: 2000 generation tokens Embedding Dimension: 112 Transformer Layers: 4 Attention Heads: 4 Head Dimension: 28 Context Length: 128 tokens FFN Hidden Dimension: 448 Default Max Tokens: 32 Default Temperature: 0.8 Default Top-p: 0.9 Default Top-k: 100 Frequency Penalty: 0.5 Quantization: 4-bit signed nibbles