Researchers introduce PACT, a system for LoRA fine-tuning on consumer GPUs that uses native NVFP4 matrix operations for frozen projections. The paper also addresses memory spent on activations and the cost of quantizing and dequantizing data.
EdgeKernelBench studies whether large language models can generate kernels for varied edge-computing hardware. The preprint introduces a common intermediate representation and a hierarchical benchmark that connects workloads to different backends.
A preprint introduces Koi, a cluster planner for large language model inference across heterogeneous multi-cloud fleets. It considers accelerator choice, parallelization, serving engines, replica counts and changing workloads as interdependent operating decisions.
A research preprint introduces Muse Glimmer, an open-source 30-billion-parameter causal language model with a perception encoder for agent tasks on consumer hardware. The authors describe text and image input, multi-step reasoning, tool use and failure recovery.
The preprint examines whether smartphone neural processing units can help fine-tune language models locally with low-rank adaptation. It focuses on the repeated operations involving frozen model weights that remain costly even when only a small number of parameters are trained.
The authors examine whether the matrix-vector processor inside a commodity EFR32MG24 microcontroller can accelerate sparse spiking-neural-network inference beyond an optimized Cortex-M33 CPU. Their toolchain maps snnTorch models to C and reorders neurons to improve accelerator use.
SlideFiLM feeds whole-slide context into tissue-image tiles to segment kidney and vascular structures. Supplementary tables report results when using just 20% of labeled training data.
Thirty-two Emirati specialists in the National Experts Programme’s artificial intelligence track are working on projects addressing national priorities across sectors. Final presentations are scheduled for January 2027. Panels of 20 leaders, programme alumni and investors will assess the projects.
The University of North Texas says its new project phase will test candidate materials for infrared sensors and examine whether they can be made with existing semiconductor processes.
Source: University of North Texas College of Science
Google Analytics can measure visits and site activity using cookies if you allow it. You can change your choice later.
Measurement and privacy details
Google Analytics measurement starts only if you allow it. After consent, automatic measurement may cover scrolls, outbound link clicks, downloads, video engagement, site searches and basic form interactions; search terms may be detected from URL parameters. This site’s own code does not send form field values; email redaction in automatic measurement is best-effort, not guaranteed. Our manual page-view event uses a cleaned URL. You can change your choice later. Visits that begin with consent may also send numeric LCP, INP and CLS performance measurements after Analytics loads. A first or later grant starts this on the next page load; withdrawing consent stops these performance reports for this page. The preference cookie lasts six months; advertising measurement and personalization stay off.
Your browser is sending a Global Privacy Control signal. Analytics is off by default; you may explicitly change this choice here.