Open-Tabletop-GM brings TTRPG campaigns to local models

This is a sensible way to make smaller local models usable for actual play: stop asking them to be a rules engine and let them be a narrator with memory.

• –The architecture removes the failure modes that usually sink AI GM experiments: arithmetic, turn order, and state tracking are deterministic and persisted in files.
• –The portable `system.md` layer makes this more interesting than a D&D toy; it is designed to adapt to PF2e, VtM, Cyberpunk RED, and other systems with different resolution mechanics.
• –The LM Studio guidance is a real signal that the project has moved beyond vibes into tuning for laptop-class hardware, not just high-end cloud models.
• –The optional Flask/SSE display companion is a practical add-on, especially for couch co-op or TV play, because it externalizes the campaign state instead of burying it in the chat.
• –The main risk is still model quality on narration and script-call discipline at smaller sizes, but the repo’s own guidance suggests that 14B-class models are already in the viable zone.