Open Dungeon runs Gemma 4 + FLUX locally with 256K context under 8GB RAM
Private AI Dungeon alternative fits 12B model and full story history on a single machine.
A developer known as akroletsgo has released Open Dungeon, a fully local alternative to AI Dungeon that prioritizes privacy and performance. The project combines Google's Gemma 4 12B model (quantized with QAT to Q4) running through Ollama for narrative generation, and FLUX for on-the-fly image creation whenever a scene warrants a visual. All computation happens on-device—no API keys, no cloud uploads. The standout technical feat is that Gemma 4's architecture barely grows its KV cache at long contexts, allowing the 12B model to operate at its full 256K token window while peaking at roughly 7.7GB of RAM. This means the AI can retain the entire story history without memory blowup.
Open Dungeon offers the standard interactive fiction controls: Do, Say, and Story modes, plus Continue, Retry, Erase, and the ability to edit any line. As the context window fills, old scenes are summarized into a running digest so the narrator never forgets earlier chapters. The UI lets users pick their model and shows the estimated RAM cost upfront. Builds are available for Mac via one-click release, or users can run from source. The project is MIT-licensed and the developer invites the community to break it and suggest missing features. The GitHub repository is https://github.com/newideas99/open-dungeon.
- Runs Gemma 4 QAT (Q4) 12B at full 256K context using only ~7.7GB RAM due to efficient KV cache.
- Generates images locally with FLUX on applicable scenes; no internet or API keys needed.
- Includes Do/Say/Story modes, line editing, and automatic summarization of old scenes to preserve story continuity.
Why It Matters
Enables private, high-quality AI roleplay with long memory on consumer hardware without cloud dependency.