Research
65 updates on Research.
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.
Amazon survey puts some small models at 10 to 100 times their parameter count
A survey of about 160 papers finds 1B to 8B models matching much larger ones on single benchmarks, and shows where models below 1B fall apart.
BlueLM-V-3B runs a multimodal model on a Dimensity 9300 NPU in 2.2 GB
CUHK and vivo AI Lab report 24.4 tok/s and 2.2 GB peak memory for a 3B vision-language model on a MediaTek Dimensity 9300.
PhoneLM searches for a fast architecture before training it and hits 58 tok/s
BUPT researchers picked their 0.5B and 1.5B transformer shapes by measuring speed on a Snapdragon 8 Gen 3 first, then pre-training the winner.
Apple team finds H100 last on tokens per dollar for models up to 2B
Seven Apple authors measured tokens per dollar for LLaMA-style models from 100M to 2B and found H100s last at every size, behind cheaper A100s.
CoMiGS splits on-device fine-tuning into shared generalists and private specialists
EPFL researchers split LoRA adapters into shared generalists and device-local specialists, and report 47.19 test perplexity against 54.38 for training alone.
ElastiLM resizes a shared phone LLM per request and switches in 0.31 seconds
Peking University and BUPT built an on-device LLM service that fits model and prompt to each request, switching sub-models in 0.31 seconds.
Ai2 releases OLMoE, 7B parameters with 1B active per token
Ai2's mixture-of-experts model holds 6.9B parameters but runs 1.3B per token, and its iOS app runs it offline on an iPhone 15 Pro or newer.
Spectra ships ternary 3.9B models that match 4-bit quantisation at half the bits
Nolano AI trained 54 models from 99M to 3.9B parameters on the same 300B tokens in ternary, quantised and half-precision form to compare them by bit size.