On-device AI models for Android
OfflineMate runs quantized language models on your phone instead of in the cloud. Here’s every model it offers, how large each download is, and how much RAM each tier targets.
How to choose a model
Start with your phone’s RAM, because the model has to fit in memory alongside Android and your other apps. Pick the tier that matches, then use its default model. It’s the option OfflineMate suggests for that tier.
A larger model generally gives better answers, but it takes more storage and memory. If a model fails to load, the app can fall back to a lighter tier.
Want to ask about photos? Pick a vision model: LFM 2.5 VL 450M on Lite, LFM 2.5 VL 1.6B on Standard or Full, or Gemma 4 E2B multimodal on Full. They have an image icon in the app’s model list. The Gemma 4 E2B default on Standard reads text only; its multimodal build is a separate, larger download because it includes the vision and audio encoders.
Lite tier: 3–4 GB of RAM
For budget phones. Lite skips semantic notes search to stay fast.
| Model | Size | Role |
|---|---|---|
| SmolLM2 135M | ~534 MB | Default |
| SmolLM2 360M | ~1.3 GB | Alternative |
| LFM 2.5 350M | ~432 MB | Alternative |
| LFM 2.5 VL 450M | ~619 MB | Alternative |
Standard tier: 6–8 GB of RAM
For most recent phones. Includes semantic search over your notes.
| Model | Size | Role |
|---|---|---|
| Gemma 4 E2B | ~2.5 GB | Default |
| Llama 3.2 1B | ~1.1 GB | Alternative |
| LFM 2.5 VL 1.6B | ~2.3 GB | Alternative |
| Qwen 3 1.7B | ~1.2 GB | Alternative |
| SmolLM2 1.7B | ~970 MB | Alternative |
| Qwen 3.5 2B | ~2.9 GB | Alternative |
| MiniCPM5 2B | ~1.3 GB | Alternative |
Full tier: 10–12+ GB of RAM
For flagship phones. Uses the largest models OfflineMate offers.
| Model | Size | Role |
|---|---|---|
| Qwen 3 4B | ~2.6 GB | Default |
| Llama 3.2 3B | ~2.4 GB | Alternative |
| LFM 2.5 VL 1.6B | ~2.3 GB | Alternative |
| Gemma 4 E2B multimodal | ~4.2 GB | Alternative |
| LFM 2.5 2.6B | ~1.7 GB | Alternative |
What the sizes mean
The sizes above are estimates for the chat model file alone. Setup also downloads speech and search files, so the total is higher: about 430 MB–1.3 GB for Lite, and about 2.5 GB for Standard and 2.6 GB for Full with their default models.
The “135M” or “4B” in a model’s name is its parameter count, in millions or billions. These builds are quantized, which shrinks them enough to run on a phone.
That’s why some Standard models, like Llama 3.2 1B or SmolLM2 1.7B, download at about 1 GB, less than some Lite files. The parameter count, not the download size, sets how much work the phone does for each word. A 1–2B model is many times more work than a Lite model. On a 3–4 GB phone it would answer very slowly or be closed by the system, so these models are in Standard, for phones with 6 GB of RAM or more.
OfflineMate downloads these models from Hugging Face. Select a model name in the tables above to see its source page.
New to local AI? Read how to run AI offline on Android.