Stop routing docstrings to 70B models with on-device AI on Snapdragon | BRKSP90

This session from Microsoft Build 2026 demonstrates how to build a three-tier inference routing architecture using Snapdragon X2 Elite’s 80 TOPS NPU to route simple tasks like docstrings to on-device models (≤13B), medium tasks to on-prem (14B-34B), and complex tasks to cloud (70B+). Developers will learn quantization trade-offs, routing logic based on entropy and logic depth, and how to achieve 67% token reduction and 70% latency improvement while keeping most code local.