ipads-sai/mobiinfer-android is a C++ Android repository that runs 35B-A3B-class MoE models on device on the Qualcomm Hexagon NPU of Snapdragon phones. The latest release is v0.1-alpha, and the provided information contains no specific list of changes for that version. It is published under the MIT license.
The implementation keeps expert weights in flash and streams in the data it needs, so the whole model is never loaded into memory. LES (lookahead scheduling) plans in advance what will be read next to hide storage latency, and it applies DSpark speculation. According to the repository description, it is based on a Hexagon NPU backend taken from llama.cpp-omni and includes a vector-unit tensor core, flash attention, a gated delta net, direct memory access queues, and multi-device partitioning. Per-layer mixed-precision quantization and reshuffling of experts into contiguous storage are configured to reduce scattered reads from storage. The installable form is the Android app included in the repository.
That said, the provided information contains no performance figures or benchmarks, so the actual inference speed and the extent to which the storage bottleneck is eased cannot be judged. What v0.1-alpha presents is the execution structure and feature set on the Snapdragon NPU, and no numbers demonstrating performance have been confirmed yet.