LLM Quantization Studies

ARTMECHO.

Interactive worlds generated by quantized large language models — arcade games, sandboxes, and function labs, each one the fingerprint of a specific model and precision.

View Experiments

About

Quantization, Visualized

Large language models are typically compressed (quantized) before they run locally, trading raw precision for speed and memory. Those benchmarks are numbers in a table.

ArtMecho takes the other view: give several model-and-quant pairs the same creative prompt and look at what they build. The difference between a BF16 reference and a 3-bit quant becomes something you can see and play.

Every demo on this site is a self-contained HTML file produced by the model in a single pass — no human edits. If a game is imperfect, that is the data.

The full set of runs, grouped by experiment, is on the experiments page.