Offline models (on-device AI)
Download CX Lite, CX Balanced or CX Smart and chat without internet.
Offline models run fully on your phone after a one-time download. No internet, no tokens, nothing leaves the device. Open Settings → AI models → Providers → Offline models (also reachable from the Offline card in the model picker).
Available models#
| Name | Based on | Download | Needs | Notes |
|---|---|---|---|---|
| CX Lite | Qwen3 0.6B (int4) | ~347 MB | 3 GB+ RAM | Fastest. Quick answers, rewriting, short chats on almost any phone. |
| CX Balanced | Qwen2.5 1.5B Instruct | ~1.6 GB | 4 GB+ RAM | Best mix of speed and quality for everyday questions, notes and drafts. |
| CX Smart | Gemma 4 E2B | ~2.6 GB | 6 GB+ RAM | Smartest offline option with good Hindi and English; slower, needs a newer phone. |
Models are downloaded from Hugging Face (litert-community) and run with Google's LiteRT engine.
Using them#
- Tap Download on a model. Download on Wi-Fi only is on by default; turn it off to use mobile data.
- A notification ("Offline models" channel) tells you when the download finishes.
- Tap Use in chats, or pick the model from the model picker under Offline.
If your phone does not have enough memory CX says so and suggests a smaller model. Remove model deletes the download and frees the space. The page shows how much storage models use.
Limits#
- Offline models have a small context (roughly 1,800 to 7,000 characters) and short replies (384 to 768 tokens).
- They do not use online tools like web search.
- Quality is lower than CX AI or big cloud models; use them for quick or private tasks.
Tip
Hinglish tip: Flight mode mein bhi chat karni ho to CX Lite download kar lo, sirf 350 MB hai.