Simon Edwardsson · 2026-10-21 · notable
WorldTUI — a 360M model that pretends to be a terminal
WorldTUI is a 360M-parameter model, fine-tuned from Qwen2.5-Coder-0.5B, that predicts how a terminal screen changes after each keypress. There is no real shell. It runs in the browser on WebGPU and can fake vim, tmux and snake.
A tiny language model that imagines a whole terminal session, one keypress at a time, with no shell behind it.
Key specs
| Parameters | ~360M |
|---|---|
| Shell bench | 80 / 106 |
What is it?
WorldTUI is a demo by Simon Edwardsson of a language model acting as a terminal. Each keypress goes in, and the model predicts the next screen, so it can appear to create files, edit in vim, use tmux and play snake without any real filesystem. The model has about 360M parameters after most of the Qwen2.5-Coder-0.5B embedding table was dropped, and an int4 build is a 208 MB download.
How does it work?
The screen is treated as a grid of characters, and the model outputs row updates plus the cursor position instead of full frames, with a COPY operation for cheap scrolling. Rows are decoded in parallel, then speculative decoding fixes mismatches between them. Training used about 336k synthetic episodes (around 2B tokens) recorded from real programs in Docker, plus a DAgger-style method that teaches the model to repair its own mistakes.
Why does it matter?
On the author's ShellBench the fine-tuned model scores 80 of 106, against 60 for a prompted Qwen3-0.6B, which shows how far a small model can go when trained on one narrow world. It is a hands-on example of a text 'world model' that runs locally in a browser tab, though the author notes weak filesystem consistency and Python support, and the weights are not public yet.
Who is it for?
ML hobbyists and world-model researchers
Try it
https://simedw.com/projects/worldtui (desktop Chrome or Edge with WebGPU)