OffRoad runs powerful open models like Gemma, Qwen and Phi directly on your phone. Your conversations never leave it, and it keeps working when the signal doesn't.
Open models from
Everything you expect from an AI chat app, running on the hardware in your pocket.
Replies stream in as the model thinks, right on your device. Stop, regenerate or copy any answer, and pick up past conversations where you left off.
Drafts emails, brainstorms ideas and writes code with proper formatting: headings, lists, tables and code blocks with one-tap copy.
Browse 14 open models with clear sizes, RAM needs and capabilities. Download with progress and cancel, and delete any time to free space.
Snap a photo or pick one from your library, and vision models describe, read or explain it.
Dictate messages or use hands-free voice mode with replies read aloud by your phone’s offline voice.
Turn on reasoning to let supported models think step by step, and peek at how they got there.
Run on your phone’s GPU for faster replies, or CPU for maximum compatibility.
Follows your system theme, or choose your own.
Delete a message, a chat or everything at once. It’s your data.
OffRoad was built so there’s nothing to leak: no servers to breach, no accounts to hack, nothing to sell.
Prompts, images and recordings are processed by the model on your phone. Never sent to a server.
Chats live in a database inside the app. Delete them anytime, or uninstall to remove everything.
The network is used to fetch models you choose from Hugging Face. That’s it.
No sign-in, no analytics, no advertising SDKs. Not even on this website.
Free on the App Store and Google Play. No sign-up.
Start with Qwen2.5 0.5B, about 500 MB, or pick a bigger one for your phone.
On a plane, underground or off-grid. It all runs locally from here on.
Sizes and RAM guidance straight from the app’s catalogue.
| Model | Maker | Best for | Size | RAM | Capabilities |
|---|---|---|---|---|---|
| Gemma 4 E2B | Multimodal Assistant | 2.5 GB | 3 GB | VisionAudioReasoning | |
| Gemma 4 E4B | Multimodal Assistant | 3.6 GB | 6 GB | VisionAudioReasoning | |
| Gemma 3n E2B | Multimodal Assistant | 3.1 GB | 2 GB | VisionAudioHF token | |
| Gemma 3n E4B | Multimodal Assistant | 6.3 GB | 3 GB | VisionAudioHF token | |
| Gemma3 1B | General Assistant | 996 MB | 2 GB | HF token | |
| Gemma3 270M | On-Device Inference | 293 MB | 1 GB | HF token | |
| FunctionGemma 270M | Function Calling | 264 MB | 1 GB | ||
| Qwen3 0.6B | Alibaba | Reasoning Assistant | 572 MB | 1 GB | Reasoning |
| Qwen2.5 1.5B Instruct | Alibaba | General Assistant | 1.6 GB | 2 GB | Code |
| Qwen2.5 0.5B Instruct | Alibaba | General Assistant | 501 MB | 1 GB | |
| SmolLM 135M Instruct | HuggingFace | Lightweight Chat | 137 MB | 1 GB | |
| Phi-4-mini-instruct | Microsoft | Reasoning & Coding | 3.9 GB | 4 GB | ReasoningCode |
| DeepSeek R1 Distill Qwen 1.5B | DeepSeek | Reasoning | 1.8 GB | 4 GB | Reasoning |
| FastVLM 0.5B | Apple | Vision-Language | 537 MB | 2 GB | Vision |
HF token models require accepting the publisher’s licence on Hugging Face and adding a free token in Settings.
Yes. You need a connection once, to download a model from Hugging Face. After that, chatting, image questions and voice all run on your phone with no network at all, including in airplane mode.
No. The model runs on your device and conversations are saved in a database inside the app. There are no accounts, analytics or advertising SDKs. The only network traffic is the model downloads you start.
iPhones on iOS 16 or later and Android phones on Android 11 or later (64-bit). Small models like Qwen2.5 0.5B run on almost anything; larger ones such as Gemma 4 E4B want a recent flagship with 6 GB+ RAM. Each model lists its requirements in the app.
From about 140 MB (SmolLM 135M) to 6.3 GB (Gemma 3n E4B). The Storage tab shows exactly what each model uses, and you can delete one at any time; your conversations are kept.
A few publishers, including Google for some Gemma models, require you to accept their licence first. Accept it on huggingface.co, create a free read token, and paste it in Settings. It's stored in your phone's secure keychain and only ever sent to Hugging Face.
Yes, OffRoad is free to download and use. The models are open-weight releases from their publishers.
Free on iPhone and Android. No account required.
OffRoad