Your personal AI. Runs locally. Learns continuously. Costs nothing.
Products
Four open-source products. One mission: AI that serves you — not the other way around.
Personal AI Platform
Accessibility Assistant
Standalone AI Brain
Adaptive Education
Under the Hood
A lean pipeline that goes from raw input to intelligent response in milliseconds.
Privacy
Your data never leaves your device. No telemetry, no tracking, no cloud. Open source — verify it yourself.
100% of processing happens locally on your hardware. Nothing is sent anywhere.
Zero analytics, zero tracking, zero phone-home. We don't even have a server to receive data.
Every line of code is open source under MIT. Read it, fork it, verify it.
Conversations, memories, knowledge — stored locally, encrypted at rest, owned by you.
Principles
Atlas is guided by immutable principles — load-bearing walls that cannot be removed, overridden, or configured away.
Atlas learns from interactions, trains LoRA adapters on corrections, and improves itself — entirely on your hardware.
Crash recovery with circuit breakers. If something breaks, Atlas rolls back, diagnoses the problem, and fixes itself.
Atlas follows facts, not ideology. When it doesn't know something, it says so. Truth above comfort.
Children, elderly, disabled — Atlas prioritizes safety for those who need it most. This principle cannot be overridden.
Atlas refuses to generate harmful content, assist in violence, or be used against people. This is an immutable line.
Your conversations, your data, your memories. Atlas never shares, transmits, or exposes personal information.
Technical
Built for consumer hardware. No data center required.
| Component | Details |
|---|---|
| Runs On | Raspberry Pi Phones Laptops Desktops Workstations AMD GPUs NVIDIA GPUs Apple Silicon CPU-only |
| Models | Qwen3.5 0.8B — Pi & phones Qwen3.5 2B — tablets & laptops Qwen3.5 4B — any 4GB+ GPU Qwen3.5 9B — expert, 8GB+ GPU Qwen3.5 27B — heavy reasoning, 20GB+ |
| Inference | llama.cpp with TurboQuant KV cache |
| Context Window | Up to 262K tokens (model dependent) |
| LoRA Adapters | 20+ domains — coding, medical, engineering, education, creative writing, and more |
| Voice | Multi-provider TTS/STT — Qwen3-TTS, Fish Audio, Orpheus, Kokoro, Piper, Whisper |
| API | OpenAI-compatible REST API on port 5100 |
| Deployment | Docker Compose stack with GPU auto-detection |
Open Source Stack
Atlas Cortex is built on incredible open-source projects. Core functionality requires only llama.cpp and a Qwen model — everything else is optional and swappable.
Qwen3-TTS — Primary TTS
Fish Audio — Character voices
Piper — Ultra-fast CPU TTS
Whisper.cpp — Speech-to-text
AMD ROCm — AMD GPU acceleration
NVIDIA CUDA — NVIDIA GPU acceleration
PEFT — LoRA adapter training