Outward

项目介绍

Outward is an installable PWA that turns a walk into an adventure story: the narrative only moves forward when you actually go outside. GPS verifies your walking distance, then a small open-weight model running on your phone — a Gemma 3 270M distilled from a Tinker fine-tuned Qwen3.5-4B, served through wllama (llama.cpp compiled to WebAssembly) — weaves what you discovered on the walk into the next chapter. After one install it works fully offline in airplane mode: no signal needed, and nothing about your location, voice, or walks ever leaves the phone. Phase 7 of 8 is done; voice input is handled by Whisper tiny via Transformers.js. Author 452Harsh wrote in the 2026-10-11 dev.to launch post, verbatim: "I built Outward with Claude Code as my pair programmer. It helped with everything from the Tinker training loop and the eval rubric to the wllama integration." The README backs it up with full training transparency (distillation rubric 2.55 → 4.75, documented in training/RESULTS.md) and a published model on Hugging Face. Outward is the health-app anti-pattern done right: instead of gamifying step counts with badges, it makes the story itself the reward — you literally cannot progress the plot from the couch. The on-device model stack (241 MB Q4_0 GGUF cached in OPFS) also makes it a reference implementation for shipping real LLM features to phones without a server bill. MIT, pushed 2026-10-11.