[Bug]: NPU not detected on Xiaomi 14T Pro (MT6989) – Unable to select NPU or download SoC-specific models in v2.0.0

[soulafein83]

Body:

Describe the bug
After upgrading to version 2.0.0, the BOX application fails to detect the NPU on a Xiaomi 14T Pro. Because the application does not recognize the hardware, NPU/TPU acceleration cannot be selected, and consequently, the app does not trigger the automatic download of the MediaTek-specific .litertlm models (like Gemma 3n E2B or Qwen3 0.6B) as described in the documentation.

To Reproduce

Open BOX v2.0.0 on a Xiaomi 14T Pro (MediaTek Dimensity 9300+ / MT6989).

Look for NPU/TPU acceleration options or attempts to fetch dedicated neural engine models.

The NPU is completely undetected, meaning the hardware-specific models aren't even visible or available for download/testing.

Expected behavior

BOX should auto-detect the MT6989 chipset upon launch, enable the NPU/TPU selection in the accelerator options, and expose/automatically download the dedicated .litertlm model builds compiled for MediaTek Dimensity.

Environment:

Device: Xiaomi 14T Pro

SoC: MediaTek Dimensity 9300+ (MT6989)

OS: Android (HyperOS)

BOX Version: 2.0.0

Additional context

The documentation explicitly lists MediaTek Dimensity (MT6989, MT6991, MT6993) as supported hardware for the unified APK in v2.0.0. However, due to the detection failure on this device, it's impossible to even download or test the dedicated MediaTek neural engine models. Only generic GPU models remain accessible.

[jegly]

@soulafein83 Very useful, thank you, if you could capture logs that would GREATLY help, as the NPU backend stuff has been hard to smooth out and that's a focus in Box atm.