Gloopy — a tiny alien to chat with

one message at a time replies on your device

One message at a time. Gloopy answers only your last message, because he was trained on single-question chats. Earlier turns stay on your screen for you, not for him.

Internet at the start. The page downloads its small runtime once when it opens. After that, every reply is computed on your machine — no server, no uploads.

About model

What is in this file. The 8.7M-parameter model, its tokenizer and the logo are embedded here. The ONNX Runtime Web engine (wasm) is fetched once from cdn.jsdelivr.net on first load; replies then run locally in the browser.

Parameters: 8,726,016
Layers: 6 · heads 6
Hidden size: 384 · ffn 768
Context: 128 tokens
Vocabulary: 4,096 ids
Training: 10,125 steps
Examples seen: 648,000
Model file: 10.48 MB uint8
Runtime: ort-web 1.23 wasm

Validation. Best validation loss 0.4785 (perplexity 1.61) on a split that shares template families with training, so it is a formatting and persona check rather than evidence of general understanding.

Honest limit. Gloopy answers from a small authored dataset of alien conversations. Replies can be odd or miss the question, and he does not know today's world.

checkpoint …

Gloopy

Preparing model…

Meet Gloopy

A tiny alien from the moon Luma, visiting Earth to look at things and eat snacks.

Gloopy replies to one message at a time. Earlier messages stay on your screen, but are not part of the next answer.

The model runs in your browser. Its small runtime needs an internet connection when the page first opens. Replies can be odd or miss the question.

A little alien. A lot of curiosity.

Ask Gloopy about the moon Luma, his spaceship, snacks, or anything he has spotted on Earth.