What they did
The poster wanted to see how far a local model could go at game development. Over about a month they ran Qwen 3.8 Flash Next (a 4-bit Unsloth quant) on a Strix Halo mini PC under Windows, through their own fork of an inference engine and a coding harness they wrote themselves. The game is TypeScript with three.js. They worked in small slices, starting from one craft on a flat checkerboard until the physics felt right. The model also built side tools: a craft model editor, a track editor and a dashboard for training progress. Project memory lived in markdown files the harness loaded each turn, with no compaction. The one part left to run on its own was training the in-game opponents, a small neural network of about 10,000 parameters: the model ran 30-minute training rounds, scored each checkpoint on a fixed benchmark and picked the next starting point by the numbers.
How it went
The game, Hyper Clashers, shipped free on itch.io, and the poster says all of the code came from the local model with no frontier model help. The run logged 1,911 prompts, about 31,900 tool calls and 20 million output tokens against roughly 3 billion input tokens, nearly all cached. Leftover training checkpoints became different AI drivers, sorted into easy, medium and hard by benchmark speed. Debugging was slow at times, and work the model did without close guidance was weak.
Worth knowing
The poster kept every design decision for themselves and warns that the model takes shortcuts when a task is too big or vague. Turning on preserved thinking in the harness made it much faster. The model was not used for 3D art. One commenter found the renderer laggy on a laptop.