Latest in AI
WorldTUI Explained: A 360M Model That Simulates a Terminal in Your Browser
The small model predicts terminal screens without executing commands, making consistency more interesting than how convincing the output looks.
Simon Edwardsson introduced WorldTUI on October 21, 2026: a 360-million-parameter model that predicts terminal screen changes from keypresses. There is no real shell or filesystem underneath. It is an interactive experiment in learned software behaviour, rather than a new terminal for getting work done.
The distinction is worth keeping in mind before trying familiar commands. A believable response is generated output, not evidence that a program ran successfully.
How WorldTUI works
Edwardsson's technical write-up describes training on Docker terminal recordings. WorldTUI predicts changes to rows and cursor positions rather than redrawing everything. It drafts rows in parallel, then verifies them with the same model.
The useful engineering idea is to match the model's output representation to what actually changes. Reprinting a whole interface after every input wastes work on unchanged content. Predicting a compact update can make an otherwise impractical interactive experiment small enough to explore.
Where to try the demo
The WorldTUI demo identifies Qwen2.5-Coder-0.5B as the base model and runs inference locally using WebGPU. A desktop browser and physical keyboard are the intended setup, with Chrome or Edge recommended. The interface supports a shell-style starting screen and custom starting states.
Local inference describes where the model runs. It does not turn the simulated environment into your computer's operating system. Do not use its responses to validate shell commands, check calculations or establish whether files have really been written.
Why the limitations are the point
Edwardsson reports inconsistent file operations and unreliable Python results. Standalone weights were still promised for later in the launch post.
That leaves a useful question for anyone evaluating learned interfaces: does the system preserve an earlier action when the interaction becomes longer or less familiar?
A terminal makes this unusually visible. Writing one value and reading back a different one is a concrete failure, even if both screens look perfectly plausible. The same distinction is easy to miss when an AI-generated interface is judged mainly by its appearance.
WorldTUI is worth exploring as a small research demo, not adopting as infrastructure. Its more durable lesson is that the convincing part of a simulated application can arrive before the dependable part.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn engineering experiments and product launches into clear stories that developers can understand and share.