OfflineGPT brings powerful AI directly to your device — no internet, no cloud, no account required. Chat with state-of-the-art language models, generate images locally, and keep full control over your data.
Why OfflineGPT:
- Full privacy: your data never leaves your device
- Works completely offline (travel, flights, restricted networks)
- One-time purchase, optional Pro upgrade
- No subscriptions
Important to know:
- OfflineGPT is not a cloud AI replacement
- Responses are slower than cloud AI (device-dependent)
- Performance depends heavily on device hardware
- Best experience on modern phones
- If you expect instant cloud-level responses, this app is not for you
Use cases:
- Offline notes and writing
- Brainstorming and drafts
- Private AI without internet access
Device requirements:
- Android 10 minimum (Android 12+ recommended)
- Minimum 4 GB RAM (6–8 GB recommended)
- ARM64 CPU (Snapdragon 6xx / 7xx / 8xx recommended)
- Models require ~1–2 GB storage each (download once)
How it works:
- Install the app (~100 MB)
- Download an AI model (one-time)
- Chat offline – no setup, no account
Pro upgrade (optional):
- Larger memory
- Custom system prompts
- Personalization features
- Multiple models and AI tools
No subscription. One-time purchase.
Model licenses:
OfflineGPT uses open-weight models running locally on your device.
No server communication is involved.
Text models:
- LFM 2.5 – Apache 2.0
https://huggingface.co/LiquidAI
- Llama 3.2-1B-Instruct – Llama 3.2 Community License
https://www.llama.com/llama3/
- Qwen 3 Thinking Models – Apache 2.0
https://github.com/QwenLM/Qwen3
- Qwen 2.5 Models – Tongyi Qianwen License
https://github.com/QwenLM/Qwen/blob/main/LICENSE
- Gemma – Google Gemma Terms
https://ai.google.dev/gemma/terms
- Rocket-3B – CC BY-SA 4.0
https://creativecommons.org/licenses/by-sa/4.0/
Image models:
- AbsoluteReality – CreativeML Open RAIL-M
https://huggingface.co/Lykon/AbsoluteReality
- Anything v5 – CreativeML Open RAIL-M
https://huggingface.co/stablediffusionapi/anything-v5
All models are used in accordance with their licenses.
Some image models are converted to MNN format for efficient on-device inference.
Model weights are not modified beyond format conversion.