Offline-first PWA with OPFS model caching

[sauravpanda]

Context

OPFS provides 3-4x faster I/O than IndexedDB. Cache API already partially integrated in Flare's browser demo.

Proposal

• Use OPFS as primary model cache (faster than IndexedDB)
• Service worker pre-caches model on first visit
• Stream weights via fetch(), write to OPFS layer-by-layer
• Subsequent visits load from OPFS (instant start)
• Support partial caching: first N layers for fast progressive inference
• Turn any static site into an AI-powered app with zero server costs

Impact

• True offline LLM — works without internet after first model download
• PWA installable on mobile home screens