dispatchAI/tiny-model-dungeon

⚔️ Tiny Model Dungeon

An AI-powered text adventure game where the entire AI runs in your browser — no server, no cloud, no API costs.

The pitch

This whole game's AI fits in 200MB.

The dungeon master is a 360M parameter model (SmolLM2-360M-Instruct) running entirely on your device via WebGPU. Every room description, monster encounter, and treasure find is generated in real-time by a model smaller than a single photo album.

How it works

1. 1.On page load, the model (~200MB ONNX) downloads once and is cached by your browser
2. 2.WebGPU accelerates inference on your GPU (falls back to WASM on older browsers)
3. 3.Every player action generates a unique AI narration — no scripted responses
4. 4.The game never contacts a server after the initial model download

Why?

This is dispatchAI's thesis made tangible: small models are powerful enough for real applications. A 360M model generating creative dungeon descriptions proves that mobile-sized AI is not a toy — it's a product.

Tech

• —Model: SmolLM2-360M-Instruct
• —Inference: transformers.js v3.7.0
• —Backend: WebGPU (with WASM fallback)
• —Hosting: Static HuggingFace Space (free forever)
• —Cost: $0

Built by [dispatchAI](https://huggingface.co/dispatchAI) — mobile AI that moves at the speed of need.