Your hardware.
Global API.
PhoneGPU turns your iPhone and MacBook into a private, attested inference cluster. A Hugging Face Space becomes the control plane. Your devices become the neural engine.
Windsurf Bridge
A zero-config peer-to-peer data channel between your Hugging Face Space and local Apple Silicon. No open ports. No VPN. No SSH. Just scan and bridge.
The Space acts as a WebRTC signaling server. Your Mac and iPhone exchange ICE candidates, then open a direct encrypted data channel. Latency: ~5ms local, ~40ms remote.
Device Mesh
Not one device. A mesh. Your MacBook runs the large model. Your iPhone handles embeddings, vision, and privacy tasks. They hand off jobs mid-stream. One logical compute unit.
Thermal Adaptation
Jobs automatically migrate from overheating devices to cooler ones. The scheduler reads CPU temp, battery state, and thermal pressure in real time.
Model Continuity
Same model family, auto-quantized per device. Q4_K on MacBook. Q2_K on iPhone. Seamless context handoff between quantizations.
Attested Receipts
Every inference produces a cryptographically signed receipt. Model hash, device fingerprint, latency, and token count — verifiable by anyone.
Privacy Modes
Raw input for speed. Hash-only for audit. Local-only for sensitive data. The same API, three trust boundaries.
Speculative Decoding
iPhone drafts tokens. MacBook verifies. 2x throughput with the same model. The small device accelerates the large one.
Zero-Config Discovery
Bonjour + WebRTC. Your Mac finds your iPhone on the same WiFi. Your Space finds your Mac via outbound WebSocket. No IP addresses, no port forwarding.
Open the control plane
Create a session, scan the QR code with your iPhone, and your Apple hardware becomes a private Hugging Face inference endpoint.
Launch PhoneGPU