Porting Specialized Compact Models to Neural Processing Units of Personal Devices: Replicating "Cells" Across a Heterogeneous Fleet

Author: Aleksandr A. Dmitriev (ORCID 0009-0000-6164-8474), ABEL BLEO AI LAB (independent research laboratory). CC BY-NC-ND 4.0. We study porting specialized compact language models ("cells": a sub-1B base with a fused low-rank adapter serving a single function) from a GPU execution path (ollama/GGUF) to neural processing units of personal devices. In an end-to-end experiment with a production filter model (binary classification of static-analyzer findings under asymmetric error costs): (1) Core ML conversion executed on the Apple Neural Engine at f16 weight precision preserves and improves quality — recall 1.000 with 93.4% noise suppression versus 79% for the original GPU path on a 132-item holdout; (2) aggressive 4-bit LUT quantization destroys the specialized model (recall 0.455), yielding the rule "a filter cell requires >=16 bits per weight"; (3) the compiled artifact runs without reconversion across the vendor's entire device lineup, turning model growing into replicable "organ firmware" for a device fleet. A three-stack porting map (Apple ANE / Qualcomm Hexagon / AMD XDNA) is given with documented pitfalls, including vendor-disabled NPU blocks in budget CPU revisions. Version 1.0; fleet replication measurements to follow in v1.1. Third work of the cycle: continues 10.5281/zenodo.21431719 (sovereign RAG and cell growing) and 10.5281/zenodo.21440910 (compute unit selection by usage mode).