Developer Runs Village Simulation Game Using a 144M Parameter Model

On November 2, 2026, a developer posted to Reddit's LocalLLaMA community about a small village simulation powered by a 144M parameter model. In the experiment, eight villagers go through a single day, with each one assigned only one job. The developer wanted to see how far small models can go when each one is limited to a single task. The decision-making core of the system is julia-1 (144M parameters) from supersonic labs, run in ggml-org's gguf format, while the memory component uses embeddinggemma, also running on the CPU. In the post, the developer explained that the model runs on the CPU and does not generate text at all. Instead, the game world hands the model the handful of actions a villager can physically take at the moment, and the model scores them. Because the model cannot invent an action on its own, the simulation does not drift into free-form roleplay. Before each choice, the system pulls from embeddinggemma to retrieve relevant memory. The project demonstrates one approach to making small models useful for game agents without relying on text generation.