Zane-GPT

1. Pick a model

The download happens once; your browser keeps it for next time.

2. Play

Sampling settings

Temperature Low = safe and predictable, high = creative but messy. Top-k Only pick from this many likely next words. 0 = no limit. Top-p Only pick from the likeliest words covering this share of the odds. Repetition penalty Discourages repeating words. 1.00 = off. Max new tokens How long the reply can get. A token is about ¾ of a word.

How it works

Every model here was trained from scratch with PyTorch on a single AMD Radeon RX 9070 XT, using a Llama-style transformer and its own tokenizer.
The weights are squeezed to 8 bits and run by a small JavaScript engine in a background thread of this page. No server, no API.
These are tiny models (17 to 50 million parameters, about a thousandth of the big chatbots), so expect charming nonsense more than facts.