How OutsideQuest AI Operates On-Device
Most modern AI applications send every keystroke, query, and context to cloud server farms. OutsideQuest AI takes an honest, privacy-first approach using open-weight models that execute directly on your device silicon.
Open-Weight Architecture: Google Gemma 3
OutsideQuest AI integrates Google Gemma 3 (270M Instruct), an open-weight instruction model built by Google and packaged for high-efficiency browser ONNX runtime execution by the Hugging Face ONNX Community.
Why Browser-Native Open Weights Matter
Choosing browser-native open weights is a deliberate design philosophy:
Your time constraints, local landmarks, physical limitations, and sensory observations never leave your device.
No monthly credits, token costs, or paywalls. Adventure creation is completely democratized and unmetered.
Inference runs on your GPU via WebGPU or your CPU via WebAssembly inside a dedicated Web Worker.
The model, validation schemas, and prompts are open and inspectable. No closed-source black box algorithms.
Key Distinctions: Open Weights vs. Open Source vs. Browser Runtime
Open Weights: Refers to the pre-trained neural network parameters (tensors). They are distributed as static files under permissive licenses (Apache 2.0).
Open-Source Software: Refers to the OutsideQuest application code, the Web Worker orchestration, prompt generators, and Zod validation parsers.
Browser Runtime: Refers to the sandboxed execution environment (WebGPU and WASM) provided by your browser that executes the ONNX runtime safely without system-level privileges.
Honest Limitations & Offline Expectations
We believe in honest engineering without hype:
- Not Offline on First Use: The AI model weights (~200 MB) must be downloaded once over an active internet connection. Once saved to your browser's Cache Storage, it works offline.
- Instant Offline Fallback: If you are offline without the model downloaded, or wish to conserve battery and data, our verified Curated Offline Engine works instantly with zero downloads.
- Model Scope vs. Giant Cloud LLMs: A 270M parameter quantized model is tuned specifically for structured micro-adventure formatting. It is not equivalent to a multi-billion-parameter cloud cluster.
- Guaranteed Safety Filters: Because generative models can occasionally make unsupported suggestions, our strict programmatic parser automatically verifies safety rules (no eating plants, no approaching animals, staying on public paths).