On-device Models
Run language models entirely on your phone with LiteRT-LM, including offline use and background downloads.
On-device models let Mobile Agent run a language model directly on your phone's hardware with the LiteRT-LM engine. No internet connection, no remote servers, and no tokens or messages leaving your device.
How it works
Models are packaged as .litertlm files and served from the Mobile Agent on-device model catalog (on-device-models.json), which the app fetches from the repository and caches in memory for 30 minutes, falling back to the bundled copy when offline. Downloads run through a native Android WorkManager worker, so a large model can download in the background while you keep using the app.
Model files are cryptographically verified: every entry pins an exact sha256 hash and sizeBytes, so a corrupted or tampered download fails safely instead of loading.
Requirements
Each model lists a minimum RAM requirement (minRamBytes). Some are based on the model publisher's published minimum, others are derived from benchmarks or comparable models. Before enabling a model, Mobile Agent checks your device against this figure — a model that won't run well is not offered.
Recommended guide (approximate):
| RAM | Comfortable models |
|---|---|
| 6 GB | Qwen3 0.6B, LFM 2.5 1.2B, Qwen2.5 1.5B, Phi-4 Mini, DeepSeek R1 1.5B |
| 8 GB | Qwen3 4B, SmolLM3 3B, Gemini 4 E2B, Qwen3 1.7B, DeepSeek R1 1.5B |
| 12 GB | Gemma 4 E4B, Qwen3 8B, DeepSeek R1 7B |
| 16 GB | Gemma 4 12B |
Larger models also report a low-memory context window that the app falls back to when available memory is tight, so you keep functional tool use even under pressure.
Current catalog
| Model | Params | Quantization | Context | Tools | Reasoning | Min RAM |
|---|---|---|---|---|---|---|
| Gemma 4 E2B | 2.3B | mixed 2/4/8-bit | 32K | Yes | No | 8 GB |
| Gemma 4 E4B | 4.5B | mixed 4/8-bit | 32K | Yes | No | 12 GB |
| Gemma 4 12B | 12B | mixed 4/8-bit | 32K | Yes | No | 16 GB |
| Qwen3 0.6B | 0.6B | int4 | 2K | No | No | 6 GB |
| Qwen3 1.7B | 1.7B | int8 | 4K | No | No | 8 GB |
| Qwen3 4B | 4B | int4 | 2K | No | No | 8 GB |
| Qwen3 8B | 8B | int4 | 2K | Yes | Yes | 12 GB |
| Qwen2.5 1.5B Instruct | 1.5B | int8 | 4K | No | No | 6 GB |
| Phi-4 Mini | 3.8B | int8 | 4K | No | No | 6 GB |
| DeepSeek R1 Distill 1.5B | 1.5B | int8 | 4K | No | Yes | 6 GB |
| DeepSeek R1 Distill 7B | 7B | int4 block-32 | 4K | No | Yes | 12 GB |
| SmolLM3 3B | 3B | int4 block-32 | 4K | No | No | 8 GB |
| LFM 2.5 1.2B Thinking | 1.2B | int4 | 4K | No | Yes | 6 GB |
Note: The catalog changes over time as models are added. Two engines are used: CPU for the smallest models (Qwen3 0.6B, LFM 2.5) and the GPU/LiteRT runtime for the rest. The
capabilities.toolsandcapabilities.reasoningflags are shown in the UI so you know before you download whether a model can call tools or reason.
Install and use
- Open Settings → Providers → On-device.
- Tap a model card. Review its size, context window, minimum RAM, license, and capabilities.
- Download the model to your device. The worker runs in the background — you can leave the screen.
- When the download finishes, select the model in the composer's model picker.
Going fully offline
Because the runtime, tool logic, and model all live on-device, you can:
- Disable Wi-Fi and cellular data and keep chatting with an on-device model.
- Let scheduled jobs run offline (though jobs that call cloud MCP servers will report connection failures).
- Keep your conversations, memory, and workspace files private on the device.
Licensing
Each model carries its own weight license — Gemma, Apache-2.0, MIT, or LFM Open. The license is shown on the model card before download. Check the individual model pages on Hugging Face for the exact terms that apply to your use case.