What the app does.

Run open-weight models on your phone, connect your own API keys, or use our hosted models. Everything local is free and sends no telemetry.

Model catalog ↓

Offline by default

Chats, image generation and voice input run on the device. No telemetry, no ads. Nothing goes to a server unless you turn on a cloud feature.

Uses your phone's hardware

64-bit ARM, Android 8.0 or newer. Models run on the CPU, the GPU (Vulkan/OpenCL) or, on supported Snapdragon chips, the Hexagon NPU.

Free core

Local chat, image generation, Whisper, Supertonic 3 TTS, writing tools and OCR cost nothing. The subscription only covers hosted models.

Features

The main features.

Everything here runs on the device, except the Python sandbox and web search.

Vision and visible reasoning

Gemma 4 takes text, images and audio in one chat. Thinking models show their reasoning before the answer.

pcb_schematic.jpg

Identify resistor R4 and verify if 24V input causes thermal overload.

Reasoning 1.4s Gemma 4 E4B

1. R4 sits next to the LM317. Bands brown-black-red → 1 kΩ.

2. Power at 24 V: P = V²/R = 24² / 1000 = 0.576 W.

3. A typical SMD resistor is rated 0.25 W → too much.

R4 is 1 kΩ. At 24 V it dissipates about 0.58 W, more than twice its 0.25 W rating, so it will overheat. Use a 1 W resistor or drop the voltage before it.

On-device Gemma 4 E4B · offline

Python sandbox and web search

In chats with API models, the model can run Python and search the web. Needs internet.

sandbox_exec.py

import numpy as np
data = np.random.normal(50, 10, 500)
stats = {
  "mean": float(np.mean(data)),
  "std": float(np.std(data))
}
print(stats)

>>> {'mean': 49.88, 'std': 9.94}

DuckDuckGo Lite No key needed

Brave & Tavily Your own API key

Voice input and text-to-speech

Dictation with Whisper and 10 voices with Supertonic 3, both on the device.

Transcribing on device…

"Summarize today's product decisions into bullet points..."

Supertonic 3

ONNX 31 languages 10 voices Streaming

Image generation

Stable Diffusion via MNN. On supported Snapdragon chips it runs on the Hexagon NPU at roughly 0.5 s per step.

PROMPT "Alpine mountain cabin under the Milky Way, 8k cinematic lighting, hyperrealism"

Engine MNN + Hexagon NPU

Speed ~0.5 s / step (NPU)

Models 4, downloadable

Resolution 512 × 512

Generated with Stable Diffusion on Snapdragon NPU (Alpine cabin under starry sky)

512 × 512 · offline Hexagon NPU

CodeLab

Write HTML, CSS and JavaScript and see the result right away.

canvas.html shader.js ● Preview

1<canvas id="viewport"></canvas>

2<script>

3 const ctx = viewport.getContext('2d');

4 renderWaveParticles(ctx, Date.now());

5</script>

More about CodeLab

22 themes and custom system prompts

Light, Dark, Ocean, Cyber, Aurora, Matrix, Midnight Gold, 24K Gold and 14 more.

Tap a theme to preview it:

OfflineGPT On-Device

Hi!

Hello! How can I help?

Model catalog

Models you can download in the app.

Sizes are approximate. You can also import your own GGUF files or browse Hugging Face from inside the app.

Text Models (GGUF · llama.cpp)

Quantized open-weight models. Small enough for phones with limited RAM.

LFM2.5 1.2B LiquidAI

~731 MB

Llama-3.2-1B Meta

~1.4 GB

Gemma-2B Google

~1.85 GB

Qwen2.5-3B Alibaba

~2.1 GB

Rocket-3B Community

~3.1 GB

Thinking Models (Visible Reasoning)

These write out their reasoning before the answer. You can read it in the chat.

Qwen3 0.6B Thinking Alibaba

~640 MB

Qwen3 1.7B Thinking Alibaba

~1.8 GB

Qwen3 4B Thinking Alibaba

~2.6 GB

SmolLM3-3B Hugging Face

~1.9 GB

Multimodal Models (Google LiteRT)

The Gemma 4 models accept text, images and audio in the same chat.

Gemma-4-E2B-it Google

Text, image, audio ~2.58 GB

Gemma-4-E4B-it Google

Text, image, audio ~3.66 GB

Qwen2.5-1.5B-it Alibaba

Text only ~1.60 GB

Qwen3-4B-it Alibaba

Text only ~2.66 GB

Vision Understanding & Image Generation

Image description and Stable Diffusion via MNN. NPU variants need a supported Snapdragon chip.

InternVL2.5 1B OpenGVLab

4 GB RAM ~888 MB

SmolVLM2 2.2B Hugging Face

6 GB RAM ~2.48 GB

Anything V5 Stable Diffusion

Anime ~1.0 – 1.25 GB

AbsoluteReality Stable Diffusion

Photo ~1.0 – 1.25 GB

ChilloutMix Stable Diffusion

Photo ~1.0 – 1.25 GB

CuteYukiMix Stable Diffusion

Anime ~1.0 – 1.25 GB

Voice Input & Text-to-Speech

Offline dictation with whisper.cpp and text-to-speech with Supertonic 3.

Whisper Base whisper.cpp

Faster ~78 MB

Whisper Enhanced whisper.cpp

More accurate ~142 MB

Supertonic 3 ONNX

31 languages ~380 MB

Hosted Models (Cloud, subscription)

Gemma 4 31B runs at Infomaniak on renewable energy. Other models go through our gateway to the official APIs, with training disabled.

Gemma 4 31B Infomaniak

Renewable energy

GPT-5.6 Luna OpenAI API

No training

DeepSeek V3.2 & R1 DeepSeek API

No training

Explore Dedicated Model Catalog & Technical Guides →

FAQ

Common questions.

Do I need an internet connection to use OfflineGPT?

Only to download models and app updates. After that, local chat, image generation and voice input work without a connection. Web search, BYOK API mode and the Python sandbox need internet when you use them.

Is OfflineGPT free?

The core is free: local chat, image generation, Whisper voice input, writing tools, OCR and custom tools. An optional subscription (CHF 14.90/mo) adds hosted models: Gemma 4 31B at Infomaniak on renewable energy, plus API models like GPT-5.6 Luna and DeepSeek, routed through our gateway with training disabled.

What is the Python sandbox?

An isolated Python environment for chats with API models. With Power Mode on, the model can run code, call web APIs, analyze data and draw charts. It needs internet and your own API key.

What is Gemma 4 E4B?

The larger of Google's two on-device Gemma 4 models, running via LiteRT-LM. It accepts text, images and audio in the same chat and can optionally use web search.

How does image generation work locally?

Stable Diffusion runs on the device via the MNN engine. On supported Snapdragon chips, the NPU variants run on the Hexagon NPU, which is faster and uses less battery than the CPU.

How does voice input work?

Download a Whisper model in the app (Base or Enhanced). When you tap the microphone, the audio is transcribed on the phone and never uploaded.

How does privacy work across modes?

By default, prompts, chats and files stay on the device. Data only leaves the phone when you use a cloud feature: model downloads, web search, your own API keys, or the hosted models.

Get OfflineGPT

OfflineGPT on Google Play.

OfflineGPT · offlinegpt.tech · Developed by CodeAlp · com.offlinegpt.app