Cubzh and Gigax just dropped NPC-Playground, a browser-based 3D demo that puts LLM-powered characters in front of anyone curious enough to click a link. The pitch is straightforward: walk around a voxel world, talk to NPCs, and then crack open a Lua file to teach them something new. No downloads, no setup — it runs in a Hugging Face Space.
The stack is three pieces. Cubzh supplies the game engine, positioning itself as an open-source Roblox alternative with over 25,000 community-made voxel items in its library and a Lua scripting API. Gigax handles the brains, using fine-tuned models built around function calling. The flow works like this: the model reads a text description of the 3D scene plus recent events and a list of available actions — say, jump, or attack — then outputs one of those actions tied to a specific entity in the scene. That’s a meaningful step beyond the old rule-based NPCs that repeated canned lines and wandered into walls.
What stands out here is that Gigax open-sourced its inference stack on GitHub, and the models themselves are downloadable from the Hugging Face Hub. The demo runs on cloud-hosted models, but developers who want to self-host have a path. That’s the kind of move that actually builds an ecosystem, not just a tech demo. The teams also published a tutorial through Hugging Face’s ML for Games course, walking through how to clone the Space and modify cubzh.lua to add new NPC skills.
Hugging Face’s involvement is worth noting — Spaces has quietly become the default place to host interactive AI demos, and game developers are starting to use it for prototyping. The callback to Roblox is deliberate; Cubzh wants to be the open alternative, and demos like this are how you convince creators to look past the network effects of an incumbent. Whether LLM-driven NPCs actually make games more fun at scale remains an open question. Latency, cost, and unpredictability are real constraints. But as a sandbox for figuring out what works, NPC-Playground is a smart move.
