P2P on-device AI with shared knowledge adapters. No cloud, no API keys, no data leaving your device. Community knowledge spreads peer-to-peer — like word of mouth.
A full language model runs directly on your phone using Metal GPU acceleration. No internet needed. No data sent anywhere. Private by architecture.
Lightweight knowledge modules (10-50MB) snap onto the base model. Each adapter specializes the AI — local slang, seasonal food, transit tips, news.
Adapters propagate through geographic gossip over the Spore network. Discover what nearby peers know. No central server decides what's relevant.
Your training data never leaves your device. Only the weight deltas (adapters) propagate. The gossip protocol IS the aggregation layer.
LoRAs are trained on local knowledge and shared geographically. Your AI knows what matters where you are — not a generic global average.
Load multiple adapters simultaneously. The app dynamically activates the most relevant ones for each question. Knowledge compounds.
On first launch, a ~1GB language model is downloaded once. After that, everything runs locally — even offline.
Chat with the AI. It runs entirely on your phone's GPU. Fast, private, no API calls.
When you ask a question, Mycelium finds relevant LoRA adapters from nearby peers and the Spore network. They download automatically.
Adapters activate instantly. Your small local model now has specialized knowledge — regional, topical, up-to-date.
LoRAs make a small model dramatically more useful in local context.
...is better than a big model that knows nothing about your world.
Trained by people in your city. The model speaks like a local.
Knows what's in season at your farmer's market.
Knows the unofficial tips for your bus system.
Daily-trained on local sources. Updated every morning.
Mycelium uses the Spore protocol for peer discovery, encrypted transport, and geographic gossip. Same Ed25519 identity. Same P2P infrastructure. No central server sees your data.
Learn about the protocol →