Research
65 updates on Research.
Online-SDFT reports 70.28% routing accuracy and updates a LoRA adapter on the phone
I-Ju Lin and Zhang-Wei Hong measure 70.28% accuracy against 52.78% for the best baseline, and run the LoRA update itself on an Android phone.
FBLayout fine-tunes transformers on phone GPUs 2.2 to 5.7 times faster
A MobiSys 2026 paper fine-tunes seven transformer models on phone GPUs 2.2 to 5.7 times faster than MNN, TFLite and TVM, with 4.2 times fewer cache misses.
llada.cpp runs a diffusion LLM on a Snapdragon NPU up to 42 times faster
Tsinghua and Beihang researchers report LLaDA-8B generating 128 tokens 17 to 42 times faster on a Hexagon NPU than on the phone CPU.
CAPED redacts phone screenshots before a cloud GUI agent sees them
Researchers at CUHK, Southeast University and Shandong University report a phone-side screenshot filter that cuts seeded leakage from 0.766 to 0.268.
Quant.npu runs 4-bit LLMs up to 15.1 percent faster on a Qualcomm NPU
Eight authors report a fully static integer quantisation method that runs 4-bit models up to 15.1 percent faster on a Qualcomm SM8650 NPU.
Tempo compresses hour-long video with a 2B vision model so a 4B LLM can answer
Meta AI and KAUST report 52.7 on LVBench for a 6B system in which a 2B vision-language model squeezes long video down to about 3 tokens per frame.
iPhone 16 Pro settles 41.5 percent below peak over 20 back-to-back prompts
Four authors ran Qwen 2.5 1.5B 20 times in a row on four edge platforms and report the iPhone 16 Pro settling 41.5 percent below its peak throughput.
Meta's MobileLLM-Flash goes shallow and wide for 1.8 times faster prefill
Meta designed 350M, 650M and 1.4B models by measuring latency on a Galaxy S25, reversing the deep-and-thin rule of the first MobileLLM.
FlexServe runs on-device LLM inference inside TrustZone for 4.41 percent more latency
Shanghai Jiao Tong University researchers kept model weights and prompts inside ARM TrustZone at 4.41 percent more time to first token.
ClawMobile tries system commands before screen taps and finishes all six test tasks
An agent runtime that runs on the phone itself routes tasks to ADB commands before UI taps, and the authors report 100 percent completion on six tasks.
Nemotron-Flash-1B decodes 1.9 times faster than Qwen3-0.6B on an H100
Nvidia designed a hybrid 1B and 3B model family around measured decoding latency on an H100 rather than parameter count, and released three checkpoints.
Intelligence per watt puts local model coverage at 88.7% of real queries
Stanford and Together AI measured 20+ local models on 1M real queries and propose accuracy per watt as the metric for local inference.