Beacon
When disaster cuts the signal, the help is already in the room. Beacon turns nearby phones into one offline AI cluster — sharding an LLM across their NPUs — so people get expert-level help anywhere, with no network and no setup.
Core trick
Shards one 3B LLM's layers across phones on a hotspot, sending activations between devices over TCP with no internet.
Description
$25
README excerpt
# Beacon **Distributed On-Device AI for Underserved Communities** Qualcomm × Meta ExecuTorch Hackathon · June 27–28, 2026 Beacon runs a single 3-billion-parameter language model distributed across multiple Android phones connected on a local WiFi hotspot — no internet, no cloud, no server. Each phone runs its assigned layer group via ExecuTorch on the Snapdragon Hexagon NPU. Intermediate activations transit between devices over TCP on a phone-hosted hotspot. The result is a capable AI assistant for first-aid guidance and language interpretation in refugee camps, rural clinics, and disaster zones where connectivity is unavailable. --- ## Problem Statement Three barriers cut AI off from the people who need it most: 1. **No internet connectivity** — disaster zones and remote clinics have no reliable network 2. **Language barriers** — aid workers and patients often do not share a common language 3. **Hardware constraints** — modern LLMs require high-end hardware most communities cannot access Beacon addresses all three: offline operation, multilingual voice pipeline, and distributed compute across low-cost devices. --- ## Technical Architecture ### Device Topology | Device | Role | Responsibilities | |--------|------|------------------| | **Phone A** (S25 Ultra loaner) | Coordinator | Hosts WiFi hotspot, runs App UI, Whisper ASR on DSP, LLM layers 1–16 on NPU, TTS on GPU | | **Phone B** (team device) | Worker | Connects to hotspot, runs LLM layers 17–32 on NPU, retur