What's actually running on your phone.
"On-device AI" in Ovels isn't one model — it's a small roster, chosen automatically based on your phone's hardware, with a cloud fallback behind all of it.
| Tier | Model family | Role |
|---|---|---|
| Mini | Qwen 3 · 0.6B | Fastest replies, lowest memory footprint — the default on older or storage-constrained devices |
| Mid | Llama 3.2 · 1B | A middle ground between the mini and full coaching models |
| Full | Qwen 3 · 1.7B | The most capable on-device coach, for newer hardware |
| Vision | LFM 2.5-VL · 450M | Names a dish from a meal photo for offline photo-mode recognition |
Each tier isn't a single file — it's compiled for whichever accelerator your phone actually has, chosen automatically:
The general-purpose backend, works on any supported device as the baseline.
Routes inference through the iPhone's Neural Engine for faster, more efficient replies.
Routes inference through the phone's dedicated NPU on supported Qualcomm Android hardware.
Advanced users can also upload a custom compatible model file rather than use one of the bundled tiers — treated as its own tier alongside Mini, Mid, and Full.
If no on-device model has finished downloading yet, the app doesn't block — it runs on a basic, rules-only mode until a tier is ready, then upgrades automatically. See AI Coach for how the active tier is used.