The Last Candle: a Día de los Muertos game where the local model may speak for the dead, but never decide who they were
Description and use
The Last Candle uses a local Qwen 3.5 0.8B model to classify player messages and pick from pre‑written replies, ensuring the model never reveals the spirit’s identity.
Problem
Design an offline narrative game that uses a local LLM to generate a spirit’s dialogue while ensuring the model never reveals the spirit’s identity.
Outcome
The game runs offline on macOS, Windows, and Linux, uses <1 GB RAM, and guarantees that the LLM never outputs the spirit’s identity, meeting the challenge’s constraints.
Workflow
- 1. User launches the game and optionally downloads the 1.3 GB model.
- 2. The game starts an Electron shell with a Nuxt SPA and a FastAPI loopback service.
- 3. On each turn, the user writes a message about an object on the altar.
- 4. The model classifies whether the message connects to the object (JSON boolean).
- 5. The game builds a list of allowed replies from authored content based on object, recovery state, language, and classification.
- 6. The model selects one reply ID from the list and the game displays the corresponding pre‑written text.
- 7. The game tracks time; if the spirit remembers enough before dawn, she leaves peacefully; otherwise she fades lost.
Evidence
An open-weight model runs entirely on your computer and does exactly two things: it listens to what you write, and it decides what she says back. It is not allowed to decide who she was.
The model never writes a word the player sees. Two calls per turn: 1. Classify. 2. Select.
The runtime model is Qwen 3.5 0.8B, Apache 2.0, chosen by a benchmark I will get to.
Installers for macOS (Apple silicon), Windows and Linux are on the v1.0.0 release. They are between 160 and 290 MB and do not include model weights.
- Record date: 2026-11-01T03:29:05.405Z