I built a small browser incremental game where you start by answering tasks at your desk to save up to rent a GPU, and hire a real Qwen3-0.6B model running locally in your browser over WebGPU (~358 MB download).
Throughout the game you buy upgrades to your model's harness, things like JSON structured output or date tools. This actually changes how the model behaves. You can even write your own prompts.
It uses MLC WebLLM on Chrome and a custom WebGPU inference engine and BPE tokenizer on Firefox (because shader buffer limitations)
Throughout the game you buy upgrades to your model's harness, things like JSON structured output or date tools. This actually changes how the model behaves. You can even write your own prompts.
It uses MLC WebLLM on Chrome and a custom WebGPU inference engine and BPE tokenizer on Firefox (because shader buffer limitations)
I also did a write-up here: https://simedw.com/2026/10/27/harness-corp/