Adding a Model
Choose between on-device and cloud models and connect a provider in Settings.
Mobile Agent supports two fundamentally different kinds of models:
- On-device models that run directly on your phone's hardware and need no internet connection.
- Cloud models that run on a provider's servers — you connect the provider once and pick its models.
You can mix both freely. Each conversation picks its own model from Settings → Providers or the model picker in the composer.
On-device models
On-device models are downloaded from the Mobile Agent model catalog and run with the LiteRT-LM engine. They are ideal for privacy-sensitive work, offline use, and fast turnarounds — but they are smaller, so choose the size that matches your device's RAM.
The catalog currently includes Gemma 4 E2B/E4B/12B, Qwen3 0.6B–8B, Qwen2.5 1.5B, DeepSeek R1 Distill 1.5B/7B, SmolLM3 3B, LFM 2.5, and Phi-4 Mini. To install one:
- Open Settings → Providers → On-device.
- Tap a model to see its size, context window, minimum RAM, license, and tool support.
- Tap Download. The model downloads in the background — you can keep using the app.
- Once complete, select the model in the composer.
See On-device Models for the full list and requirements.
Cloud providers
Cloud models run on the provider's servers and support the largest, most capable models. Mobile Agent connects to them on your behalf using your own keys or OAuth sign-in.
Supported providers:
| Provider | Auth | Notes |
|---|---|---|
| OpenAI | OAuth or API key | GPT models, image generation, Responses API |
| Anthropic | API key | Claude models with reasoning |
| OAuth | Gemini models | |
| xAI | API key | Grok models |
| OpenRouter | API key | Many models through one gateway |
| Ollama | none | Your local Ollama server, e.g. http://localhost:11434 |
| OpenAI-compatible | API key | Any /v1/chat/completions or Responses-compatible endpoint |
To add a provider:
- Open Settings → Providers.
- Choose a provider from the list.
- Complete OAuth sign-in, paste an API key, or enter a custom base URL.
- Pick a default model preset (or let the app discover available models).
Security: API keys are stored in the app's encrypted on-device secret store and are only used to talk to the provider you configured. They are never uploaded to Mobile Agent servers and never sent to the model.
Reasoning effort
Models with reasoning support expose a reasoning effort level from none through minimal, low, medium, high, and xhigh. Higher effort produces deeper reasoning for hard problems at the cost of latency and tokens; lower effort is best for quick, routine replies.
Custom endpoints
For OpenAI-compatible providers, enter the full base URL (for example https://your-gateway.example.com/v1). Custom profiles support both the Chat Completions and Responses transports, letting you point Mobile Agent at local servers, gateways, or self-hosted stacks.
Choosing between on-device and cloud
| On-device | Cloud | |
|---|---|---|
| Internet | Not required | Required |
| Privacy | Data never leaves the phone | Messages go to the provider |
| Model size | Small to mid (0.6B–12B) | Up to frontier models |
| Speed | Dependent on device | Dependent on network |
| Cost | Free | Uses provider billing |
| Tool use | Only tool-capable models | Most models |
The ResolvedModel picks in the composer show which capabilities each model supports — tools, image input, image generation, and reasoning — so you can pick the right model for the task.