The compact model, WorldTUI, which has been trained using the Qwen2.5-Coder-0.5B model as its foundation, boasts a quantized size of a mere 208MB and is capable of running locally. This innovative model eliminates the need for a genuine Linux system or file system, as it adeptly forecasts terminal display content solely based on user-inputted commands. It excels at simulating routine terminal operations, such as 'ls' and 'cd', and even goes as far as to replicate the operational effects of software like Vim, Python, and tmux, all while preserving contextual operational memory.

To fine-tune the model, developers executed a real Linux terminal within a Docker container, mimicking human operations, including intentional typing errors, and amassed a substantial volume of operational log data. Additionally, they leveraged error-correction samples to diminish the incidence of inconsistencies. Despite its susceptibility to errors when confronted with complex commands and its current limited practical applicability, WorldTUI nonetheless pioneers a fresh approach to training models that emulate software behavior.