AEON
Adaptive Evolution On-device Network
A self-evolving AI companion that lives on your Mac and iPhone, gets smarter every day, and is yours forever. No cloud. No subscription. No API calls. Just pure on-device learning.
Core Principles
Identity
One entity that grows. Not a tool. A companion.
Self-Evolution
Gets smarter from its own thinking. No external APIs.
Yours Forever
Runs on YOUR device. No subscription. No cloud.
Transparent
See what it learned, when, and why. Roll back anytime.
The Self-Evolution Cycle
↓
Evolve → Stronger → Think Again
Every scan of your device, every interaction, every question—the model learns. Overnight, it consolidates that learning into stronger reasoning and faster inference. You wake up to a smarter AEON.
The 7 Weapons
Techniques that make a tiny 3.8B model compete with giants:
Test-Time Compute Scaling
Think longer, not bigger. Monte Carlo Tree Search (MCTS) enables reasoning time scaling—the model explores more solution paths before deciding.
Chain-of-Thought Distillation
Inherit reasoning from giants. Learn step-by-step reasoning patterns from larger models, compress them into efficient LoRA adapters.
Sparse Activation (MoE)
12B knowledge, 3B compute. Mixture of Experts routes tokens to specialized adapters, paying full compute only for relevant experts.
Self-Play Evolution
Train on your own best reasoning. The model plays against itself, learning from wins and failures without external supervision.
Process Reward Models
Judge every step. Process rewards evaluate reasoning at each token, kill bad paths early, and guide the MCTS search tree.
Speculative Decoding
2-3x speed. Draft model + main model work in tandem—drafts generate candidate tokens, main model validates in parallel.
Domain-Specific LoRA Stacking
Specialized adapters per skill. Code, math, writing, reasoning—each domain gets its own learned weights that layer on top of the base model.
Architecture
Multimodal Core
Why This Matters
Runs Locally
No internet required. Your data never leaves your device. AEON lives in your control, not in someone's cloud infrastructure.
No Subscription
Pay once, own forever. The model runs on your hardware. No recurring fees, no feature paywalls, no "your limit resets tomorrow."
Gets Smarter Over Time
Every interaction trains the model. Unlike static models, AEON actively learns from your preferences, your style, your needs.
True Privacy
Zero telemetry. Zero tracking. Your learning data is encrypted on-device. You own your model's growth history.
Platform Support
🖥 Mac (M1/M2/M3)
- Full brain (training + inference)
- Sleep consolidation overnight
- LoRA adapter evolution
- Process reward training
- 8GB RAM minimum
📱 iPhone (A15+)
- Fast brain (inference only)
- CoreML quantized model
- Adapter sync from Mac
- Instant responses
- Background learning
Tech Stack
Current Status
Month 1 — Building Foundation
AEON is in active development. The core architecture is designed. Next phases include:
- MLX inference engine with Phi-4-mini loading
- MCTS reasoning loop + process reward model
- LoRA training pipeline for on-device evolution
- Sleep consolidation scheduler
- Curiosity engine for self-challenge generation
- Mac native app UI (SwiftUI)
- iPhone app with CoreML inference
- Adapter versioning + rollback system