Research
73 updates on Research.
ShadowNPU scores attention on the Snapdragon NPU and reports 4.5 times faster inference
Peking University and BUPT move attention token scoring onto a Snapdragon NPU and report up to 4.5 times faster inference using a single CPU core.
P/D-Device prefills in the cloud and decodes on the phone, cutting TTFT 60%
Huawei researchers split prefill and decoding between cloud and phone, reporting time to first token down at least 60% and cloud throughput up to 15x.
Benchmark of 68 small language models finds architecture outweighs size on device
A study of 68 models from 100M to 5B puts Phi-3 near 70 percent accuracy and finds first-token time and memory tracking architecture, not parameter count.
Vijay Janapa Reddi argues edge generative AI needs models under 1B parameters
In ACM Queue, Vijay Janapa Reddi argues generative AI reaches edge devices only through small deployable models, with sub-1B ones the viable case.
Alibaba gives Qwen3's 0.6B and 1.7B models a reasoning switch
Qwen3-0.6B and Qwen3-1.7B carry the family's switch between a reasoning mode and a fast mode, with 32K context and Apache 2.0 weights.
Google's ML Drift runs Llama 3.1 8B on a phone GPU at 12.7 tokens per second
The paper measures 37.1 decode tokens per second for Gemma2 2B and 12.7 for Llama 3.1 8B on the Adreno 750 GPU of a Samsung S24.
D2MoE picks a bit-width per token, 1.39 times the throughput at up to 53 percent less memory
A MobiCom 2025 paper routes every token to an expert and a bit-width, reporting up to 1.39 times the throughput of EdgeMoE at up to 53 percent less memory.
ROMA keeps a 4-bit 3B model in on-chip ROM and reports 31,800 tok/s in synthesis
A synthesised 7 nm accelerator holds a 4-bit 3B Llama in read-only memory and the LoRA adapter in SRAM, with no chip and no FPGA prototype built.
Raspberry Pi 5 beats human reading speed only up to Phi 3.5 mini
Helsinki and EURECOM researchers measured 11 models on a Raspberry Pi 5 and a Jetson Orin Nano, where memory and battery gave out before speed did.
MobiLLM fine-tunes OPT-1.3B in 4.5 GB on device by moving backpropagation to a server
Researchers fine-tuned OPT-1.3B on a Jetson Xavier NX in 4.5 GB by keeping the frozen model on the device and the trainable side network on a server.
GenAI at the edge survey lists 12 accelerators, 8 of them only simulated
A Johns Hopkins and Duke survey of generative AI on edge devices puts peak accelerator efficiency at 74.34 TOPS/W, with 8 of 12 designs only simulated.
Meta's ParetoQ compares five bit widths and puts the accuracy cliff at 1 bit
Meta trained the same models at 1, 1.58, 2, 3 and 4 bits under one recipe and reports 1.58-bit to 3-bit beating 4-bit at matched memory.