XaeroAI

Nano AI models for peer-to-peer distributed learning and inference.

What is this?

XaeroAI provides small, efficient AI models that can:

• Run inference on mobile devices
• Learn from each other over a network
• Coordinate tasks without central servers

Each model is a network peer that discovers other models and collaborates on AI tasks.

Models

We provide 4 nano models, each optimized for specific tasks:

Model	Size	Purpose
Phi-2 Code	63MB	Code completion and analysis
CodeGen 350M	60MB	Code generation
TrOCR Small	56MB	Handwritten text recognition
TrOCR Base	47MB	Printed text recognition

Total: ~226MB for all models.

How it works

Model Storage

Models are compressed and split into 25MB chunks for efficient distribution:

models/nano/
├── phi2_ultra_code_nano_part_000     # 25MB chunk
├── phi2_ultra_code_nano_part_001     # 25MB chunk  
├── phi2_ultra_code_nano_manifest.txt # Assembly instructions
└── ... (other models)

Memory Loading

Models use memory-mapped loading via Candle:

• Load only the parts needed for current tasks
• Efficient memory usage on mobile devices
• Stream model weights as needed

Peer-to-Peer Learning

Models connect through XaeroFlux subjects:

• Discover other models in the network
• Share learning updates via CRDT operations
• Coordinate multi-model tasks
• Learn from workspace events (whiteboard, chat, etc.)

Architecture

AI Model Peer
├── Local inference (Candle + mmap)
├── Network coordination (XaeroFlux subjects)
├── Learning from events (CRDT operations)
└── Task collaboration (gossipsub)

Each model subscribes to relevant subjects:

• workspace/ai/discovery - Find other models
• workspace/ai/tasks - Coordinate inference requests
• workspace/ai/learning - Share parameter updates
• workspace/{id}/data - Learn from workspace activity

Usage

use xaeroai::{XaeroAIModel, XaeroAISubject};

// Load a nano model
let model = XaeroAIModel::load("phi2_ultra_code_nano")?;

// Connect to network
let subject = XaeroAISubject::new(model.id)?;
subject.subscribe_to_discovery();

// The model automatically:
// - Announces itself to other peers
// - Responds to inference requests
// - Learns from workspace events
// - Coordinates with other models

Integration

XaeroAI integrates with:

• XaeroFlux - Event streaming and CRDT operations
• XaeroID - Peer identity and authentication
• Candle - Efficient model inference
• Flutter/Dart - Mobile app integration

Building

1. Reassemble models from parts:

# Models auto-reassemble during build
cargo build

2. The build process:
   • Combines model parts into complete models
   • Sets up Candle for memory-mapped loading
   • Prepares models for peer-to-peer coordination

Development

Models are created from foundation models using:

1. Quantization (4-bit) for size reduction
2. Knowledge distillation for nano architectures
3. Compression and splitting for git-friendly storage

See scripts/ for model creation and training tools.

License

MPL-2.0