Tian AI: I Built an AI Assistant That Runs 100% Offline on My Phone
I got tired of paying $20/month for ChatGPT, sending my private conversations to servers I don't control, and being useless without internet. So I built my own AI that runs entirely on my phone.
The Problem
Every mainstream AI has the same three problems:
- Your data leaves your device — Every query goes to someone else's server
- Subscription fees — $10-200/month, forever
- No offline mode — Useless when you have no signal
I wanted something that works like Jarvis from Iron Man — a private AI that lives on my device, knows my data, and works anywhere.
What I Built: Tian AI
Tian AI is an open-source, self-evolving AI system that runs completely offline on Android (via Termux), Linux, or any device that can run Python.
Core Specs
| Feature | Detail |
|---|---|
| LLM Engine | Qwen2.5-1.5B via llama.cpp (runs on ARM/CPU) |
| Project Size | 770+ Python files, 171K+ lines of code |
| Knowledge Base | SQLite with millions of indexed concepts |
| Backend | Flask REST API |
| Privacy | Zero data leaves your device |
| Cost | Free & open source |
| GitHub | github.com/3969129510/tian-ai |
Architecture
Tian AI is built around five specialized engines:
1. Thinker — Three-tier reasoning engine:
- Fast Mode: Simple responses (~1-3s on mobile)
- Chain-of-Thought: Step-by-step reasoning for complex problems
- Deep Mode: Multi-perspective analysis with reflection and synthesis
2. Talker — Multi-turn conversation with short/long-term memory
3. Knowledge Retriever — Million-entry SQLite knowledge base with 0.04-0.1s lookup time
4. Agent Scheduler — Autonomous task planning, dependency resolution, and execution
5. Self-Evolution System — The AI analyzes its own code, suggests improvements, and patches itself
The Self-Evolution Feature
This is what makes Tian AI unique. Most AI systems are static — trained once, never changed. Tian AI has an XP/leveling system where:
- Every interaction earns XP
- Level-ups unlock new capabilities
- The system uses Python AST parsing to analyze its own code
- It generates patches, validates them, and applies them automatically
- Version tracking: M1 → M1-E1 → M1-E2 → M2
Running on Phone (Real Test)
I run Tian AI on a Realme V70s (Android) via Termux:
# Start llama.cpp server
llama-server -m qwen-1.5b-q4.gguf --port 8080 -t 4 -c 2048
# Launch Tian AI
python run.py
The 1.5B model runs smoothly on mobile hardware. Knowledge retrieval takes under 100ms. Full LLM reasoning takes 1-60s depending on complexity.
Agent System in Action
Tian AI's agent scheduler can autonomously:
- Plan and execute multi-step tasks
- Resolve task dependencies (topological sorting)
- Check safety whitelists before executing commands
- Self-evaluate after each task
- Handle file operations, code analysis, and automation
Safety is built in: whitelisted directories, no dangerous commands (rm -rf, sudo), read-only by default.
Why Local AI Matters
The AI industry is obsessed with larger models and bigger clouds. But there's a quiet revolution happening on the edge:
- Apple Intelligence runs on-device
- Llama.cpp makes local inference practical
- Qwen2.5 proves small models can be remarkably capable
Tian AI is part of this movement. It proves that a personal AI doesn't need cloud infrastructure. It doesn't need a subscription. It doesn't need your data.
Get Started
git clone https://github.com/3969129510/tian-ai
cd tian-ai
pip install -r requirements.txt
# Download Qwen2.5-1.5B GGUF model
python run.py
Support the Project
Tian AI is completely free and open source. If you find it useful:
USDT (TRC-20): TNeUMpbwWFcv6v7tYHmkFkE7gC5eWzqbrs
BTC: bc1ph7qnaqkx4pkg4fmucvudlu3ydzgwnfmxy7dkv3nyl48wwa03kmnsvpc2xv
GitHub: github.com/3969129510/tian-ai
Tian AI — Your Private AI, Completely Offline.
Top comments (0)