Tiny language models, trained from scratch on a home PC by Kael Computing

Try Aqua Q1 live — right in your browser

This demo runs the real 17.9M-parameter model, quantized to 8-bit, entirely on your device via WebAssembly - no server, no API key, nothing uploaded. Works on phones too.

Hi! I'm Aqua Q1, trained locally by Kael Computing. Ask me anything, or say "make a website about dogs".
Loading model (18 MB)...

The models

ModelParametersTrained onWhat it does
Aqua Q5 (newest) 19.4M 780,000 chars Everything in Q4, plus teach it your own Q&A with a quick on-device train (add_qa.py) - no full retrain.
Aqua Q4 19.4M 744,000 chars Latest Aqua: general knowledge, math, geography and single-file website answers. The smartest Aqua yet.
Aqua Q1 17.9M 456,000 chars Chats and builds websites in HTML. You can try it above.
Aqua Q2 17.9M 647,000 chars Chat + full HTML page generator, saves pages to disk
Aqua Pilot 1 10.0M 24,900 chars Safe command helper: only runs ls, cd, pwd, mkdir (rm asks first)

How it works

Every Aqua model is a small transformer trained in PyTorch that predicts the next character. There is no big cloud API - the weights were trained locally on an NVIDIA GTX 1650. On this site, the model runs in your browser: the 8-bit quantized weights + onnxruntime's WebAssembly build are served as static files - nothing is ever sent to a server.

Source code

RepositoryDescription
github.com/leon1006965/aqua-q1 Aqua Q1 - weights, trainer and chat script
github.com/leon1006965/aqua-q2 Aqua Q2 - improved trainer, HTML page generation
github.com/leon1006965/aqua-q4 Aqua Q4 - latest: stream-trained general-knowledge AI
github.com/leon1006965/aqua-q5 Aqua Q5 - teach it your own Q&A with add_qa.py
github.com/leon1006965/aqua-pilot Aqua Pilot 1 - command helper with a security whitelist