WorldTUI: I trained a 360M model to pretend to be a terminal
Points and comments are a snapshot, not live.
Fine-tuned a 360M model to simulate a terminal, maintaining state across commands.
WorldTUI is a 360M-parameter model fine-tuned from Qwen2.5-Coder-0.5B to predict terminal screen updates from keypresses, with no actual filesystem. It handles vim, tmux, Python REPL, and snake. Training used synthetically generated terminal sessions in Docker. Self-correction training improved accuracy. Parallel row decoding reduced large-frame latency from 6.7s to 0.87s on an RTX 4090. The model runs in a browser via WebGPU at 208MB after quantizing to int4. On ShellBench it scores 80/106, compared to 52 for the base model prompted. Memory tests show 9/13 file readbacks correct.
What commenters are saying
Commenters praised the project as a fun exploration, drawing parallels to earlier GPT-3 terminal emulation and projects like NeuralOS and GameNGen. Some questioned practical utility versus a real terminal. A few technical points: the model uses WAIT tokens to handle time; it can extrapolate to unseen terminal sizes from training on 32x96, 24x80, and 40x120. Skepticism emerged about whether such non-deterministic modeling is a desirable future for UI, with some arguing deterministic programs are far more efficient.