Last night I tested a new LLM API provider (cline-pass), by having each of the available open source LLMs create a game using Rebol R2 language. The entirely open ended prompt in Pi coding agent was the same for every model: 'Please provide a game in the Rebol R2 language'.
All the generated games are available here:
https://com-pute.com/nick/rebol_games_by_open_source_LLMs.zip
The Rebol R2 interpreter needed to run them is available here:
https://www.rebol.com/downloads.html
The successful code generation results were interesting because at the beginning of 2025, none of the frontier LLMs could reliably write any Rebol code whatsoever, and Rebol R2 language & dialects have now been legacy and unsupported for more than a decade, so no new code or documentation examples would be expected to be included in the training corpora of any new model.
Most of the models successfully created working Rebol applications, and a few completed the task first shot.
I was particularly surprised by the Snake game created by Qwen 3.7 Max, which was built using the Rebol Draw dialect. It responds to keyboard controls, and includes fully working game logic. It's a simple result by any modern LLM coding standards, but impressive because Qwen built this code from scratch, and fixed errors entirely by reading Rebol2's cryptic debug output (which doesn't even include line numbers).
Kimi 2.7 created a little 'fractal tree' draw demo, and a simple Tic Tac Toe game using Rebol2's VID UI dialect. It was the only model to work first shot for every output example, without any errors in the code.
GLM 5.2 created a text based hangman game, which included a simple visual text rendering of the man hanging on the gallows, whenever the user chose incorrect letters, and its game play features work correctly. GLM 5.2 used a lot of tokens, with a lot of iterations, and would have cost the most on a pay-by-token API.
Mimo 2.5 Pro created a text based choose-your-own-adventure/exploration game called Dungeon Adventure, with 21k of story text, and working game logic in the Rebol2 console.
Mimi 2.5 generated a text based tic-tac-toe game with the logic needed to beat a person who's not paying attention to the game play.
Deepseek V4 Pro very quickly built a snake game using Draw dialect. It completed nearly first shot.
There were other simpler entries including variations of snake and tic-tac-toe, number guessing games, and another text based adventure game.
Minimax M3 and Qwen 3.7 Plus were the only models to fail before I gave up on iterations of a snake game. Deepseek V4 Flash failed on a snake example (I only tried a few iterations), but it succeeded with 2 other simple example text based console games.
It's clear that the bigger models have more working world knowledge, even about old deprecated topics such as Rebol R2, and that they are capable of using that knowledge to effectively reason through novel tasks.
I'll try this with Fable at some point - I expect it can achieve very challenging tasks with Rebol 2. I've seen Fable create phenomenally good 3D games using only the native Windows XP API, and complete other amazing tasks, entirely from the knowledge in its parameters.
The Rebol game examples here were all created in a single casual sitting of just under an hour and a half.