Your Android phone may have a neural processor that a local AI app can use. That does not mean every model or every Android device can use it successfully.
OGAM (Off Grid AI Mobile) includes an experimental NPU (Beta) option for supported Qualcomm devices running compatible local text models. You can select it in the model settings, reload the model, and check a small task before using it for regular work.
Download the app and model first. Model inference can then run on the phone without internet. NPU selection does not require sending prompts to a remote AI service.
What the NPU option actually does
The text-model option uses Qualcomm's Hexagon acceleration path. OGAM exposes it when the device check reports support. A phone being sold with an “AI chip” is not enough to establish compatibility with this particular runtime.
The setting is marked Beta for a reason. The app describes Llama- and Qwen-style models as better candidates for this path, while some models can fall back to CPU or produce invalid output. A supported device and a downloaded model are the starting point, not proof of a useful NPU result.
The option ships in OGAM 0.0.111. It is a core text-model setting, separate from the Pro voice features.
Start with a model that already works on CPU
Choose a modest local text model that fits your phone's memory. Get one normal answer with it before changing the acceleration setting.
Use a question with an easy-to-check result, such as:
Put these numbers in ascending order: 18, 4, 12, 7. Output only the sorted list.
The expected order is 4, 7, 12, 18. This is a simple check for usable output, not a complete test of model quality or speed.
Keep the same model and request when you try another backend. That makes it easier to identify a problem caused by the new setting rather than by an entirely different model.
Select NPU and reload the model
- Open Settings → Model Settings and the text-generation controls.
- Expand Advanced.
- Under Inference Backend, choose NPU (Beta) if it is available.
- Review the NPU Layers setting. Start with the app's existing value rather than forcing the highest setting.
- Unload and reload the model so the backend change takes effect.
- Send the same short request and check the answer.
If NPU is absent, the app does not expose this path for that device and configuration. Use one of the offered backends instead. LiteRT models have their own Acceleration selector with CPU and GPU options; that is not the same NPU control.
A reply does not prove the NPU handled it
OGAM can retry initialization on CPU when the accelerated path fails. That keeps a model usable, but it means a successful answer alone does not prove the NPU ran the request.
Check the app's available generation and debug information when you need to identify the actual backend. Do not treat the selected preference as a measured hardware result.
There is no single speed gain to promise across devices and models. Load time, memory use, output quality, and generation speed can change independently. Use the backend that gives useful, stable results on your phone.
When to switch back
If the model fails to load, returns corrupt text, or becomes less useful after the change, select CPU or the offered GPU option and reload it. Try a smaller compatible model before increasing memory use.
The NPU does not remove the model's memory requirements. Model weights, the conversation context, and working buffers still need room. A larger layer setting is not a general fix for an oversized model.
Image generation has a separate device-specific acceleration path and model packages. Choosing the text-model NPU backend does not switch every feature in the app to NPU.
Try the hardware your phone actually supports
Get OGAM for Android, establish one working local model, and try the NPU option if your device offers it. Keep a simple correctness check and a known working backend so you can choose based on the result on your phone.
For further actions, you may consider blocking this person and/or reporting abuse