WorldTUI Is a 360M-Parameter Model That Fakes a Linux Terminal, and Its Weights Are Not Out Yet

Simon Edwardsson, co-founder of the computer-vision company V7, said on October 21 that he fine-tuned a 360-million-parameter language model to act as a terminal emulator, producing the screen contents of a Linux shell without running any shell. The weights are not published.

In his write-up, Edwardsson said he would "post the weights soon." The post names no licence and links no repository. A 208 MB int4 build runs in the browser through WebGPU at a demo page on his site.

What the model is built from

The base is Qwen2.5-Coder-0.5B, which Alibaba's Qwen team lists on its Hugging Face card at 0.49 billion parameters under Apache-2.0, trained on 5.5 trillion tokens. Edwardsson cut most of the embedding table, taking the model from about 494 million to 360 million parameters.

He also replaced the vocabulary with 333 tokens: 256 byte tokens, 64 row tokens and a handful of special operations. The model emits the terminal screen row by row rather than as a stream of text.

Training data is 336,000 synthetic episodes, about 2 billion tokens, recorded from real programs running in Docker. Usernames, hostnames, prompts and terminal sizes were randomised, 1.2% of keystrokes carry typos, and 35% of sessions start mid-session. Edwardsson said the run took roughly 75 GPU-hours on one RTX 4090, including later refinements.

The scores are the author's own

Edwardsson built the benchmarks himself and said none of their items were in the training data. No outside party has run them.

Model	ShellBench (of 106)	File readbacks (of 13)
Qwen2.5-Coder-0.5B, prompted	52	2
Qwen3-0.6B, prompted	60	3
WorldTUI	80	9

The post also reports a second test, PresetBench, rising from 24 to 38 of 47 at int4 after three rounds of self-correction modelled on the DAgger method.

1. Qwen2.5-Coder-0.5B, prompted52 points
2. Qwen3-0.6B, prompted60 points
3. WorldTUI80 points

Source: Simon Edwardsson, simedw.com, October 21, 2026

One figure does not reconcile. The post says its memory test uses 32 worlds and 64 readbacks, while the table reports file readbacks out of 13. The post does not explain the difference.

Speed and where it breaks

A single keypress takes about 38 ms on the RTX 4090 and about 60 ms in a browser on an M4 Mac. Decoding rows in parallel cut p95 latency on large frames from 6.7 seconds to 0.87 seconds on the 4090.

The model loses track of nested directories. In one example Edwardsson gave, lines appended with echo >> colors often do not show up in a later cat colors. Python works only for simple operations, and containers have no network, so tools such as lynx are limited.

A 16-layer distilled student passed all 64 memory readbacks at about 138 MB, but scored 65 on ShellBench against 78 for the teacher. Edwardsson dropped distillation. He also reported that the browser build runs about 20 times slower in Firefox than in Chrome and said he does not know why.

For open-weight releases that did ship with weights, see our reports on DeepSeek V4.1-Flash-1017 and Meta's Muse Spark Open.

Edwardsson said he plans to train on thousands of randomly generated applications and may try diffusion decoding in place of autoregressive decoding. No date was given for the weights.