Xiaomi

Arena.ai analyzed 105 models across 44 weekly Code Arena WebDev leaderboards from December 2025 to October 2026, finding release cadence rose from about five models per month to 12-14, with top scores climbing 3x faster than in winter. By October, Xiaomi's MiMo-V2.6-Flash outscored April's top model Claude Opus 4.7 at under one-twentieth the price, with a 65-80% win rate at equal cost.

The smolmaxxing project demonstrates running language models from 260K to 16B parameters on a Xiaomi TV Box S (Amlogic S7D, ~1GB usable RAM, no root) using custom kernels, smarter memory use and weight streaming. Qwen3.5-0.8B reaches 16.8 tok/s (2.2x stock), while an 8B MoE model hits 7.2 tok/s via on-demand expert paging from flash. All claims are measured back-to-back on real hardware, offering a reproducible engineering playbook for edge-side local inference.

A developer finds that used Redmi 13C 5G and Huawei Mate 30 5G phones, costing only ¥300-500, already pack 40-50MP cameras, GPS, storage, battery and USB-C, replacing a costly Raspberry Pi plus Arducam plus GPS stack. By rooting the phone and running the Moth Rig app over USB-C to an LED controller, the author validates old smartphones as viable embedded project controllers.

Xiaomi and Peking University released a Thing Model technical whitepaper at the CCF China Software Conference, opening practices validated across 1.16 billion AIoT devices and 200+ categories to the industry. The paper proposes a four-layer thing model architecture with a Dynamic Service mechanism and pushes thing models toward AI-agent-oriented capability contracts supporting MCP and Function Calling. It also builds a five-layer security framework on Xiaomi's MiTEE trusted system for permission control and risk-tiered AI device operations.