Tag
Alex Funk unveils SparkHome, a local AI-powered smart home harness that runs voice agents, IoT control, and AAA game optimizations (via GameSpark for GB10) on a personal device instead of the cloud.
A paper proposing the Task-Aware Diffusion Refiner (TADR), a conditional diffusion model trained against a frozen downstream anomaly detector to fix imputation bias in block-missing IIoT sensor data, improving F1 by 0.338 on SMD while running at 41.4 ms per window on a single CPU core.
An updated 2.3bpw quantized checkpoint of a Qwen 27B model for 12 GB Ampere cards ties Qwen3.6 BF16 on LiveCodeBench (84.3% vs 83.9%) and cuts KLD error by ~40% versus the previous release, using the new AmperInteger+ quant format in llamAmpere.
Breakthrough Listen and the SETI Institute are deploying agentic AI on the Allen Telescope Array, powered by NVIDIA Holoscan, to autonomously steer the instrument and process terabits per second of sensor data in real time — already achieving the first real-time AI detection of a pulsar in the Crab Nebula.
Thread discussing the trade-offs of running AI locally on-device versus calling cloud APIs, covering data control, network dependencies, per-token costs, and compute requirements.
Lasm compiles Lean 4 programs ahead of time to WebAssembly, enabling formally verified web apps that deploy to Node, serverless, and edge platforms from a single code base, with 117 proved theorems in its flagship Conduit example.
Neywa Labs 隐身模式出闸，宣布打造可在用户现有设备上运行 AI 的基础设施，主打私密、离线与无限使用，目标是让全球每个人都能用上 AI。
Nvidia announced the OpenShell Reference Architecture, a framework for building, containing, observing, and enforcing safe AI agents across enterprise infrastructure and edge systems throughout the full agent lifecycle.
CoASFL 提出一种面向异步分割联邦学习的不一致性感知框架，通过动态激活合并与一致性感知模型聚合，同时缓解客户端间水平不一致和客户端-服务器垂直不一致，在80设备边缘测试床上将训练加速1.61×至3.74×并提升最高3.26%的准确率。
Nvidia 发布了面向企业级智能体的 120B 开放权重模型 Nemotron 3.5 Super VL，以及可在 RTX/DGX Spark 本地运行的两款轻量 Bolt 模型，宣称在多项智能体任务上超越 Anthropic 和 OpenAI 的闭源模型。
该论文提出ΣΔ-Sampler优化器，将二值和三值神经网络训练损失映射为高阶Ising模型能量，并通过噪声整形与退火实现高效的边缘设备训练，在CIFAR-10、ImageNet-1k和MatMul-free语言模型上验证了接近全精度的精度，同时将训练内存减少超过50%、速度提升一个数量级。
论文提出 TRACE（Transformer Residual Activation Codebook Encoding），在每个 Transformer 层的后注意力残差处放置向量量化器并预计算码本索引化的键值表示，使分布式边缘设备推理中接收方通过查表直接恢复 K/V，实现通信-计算重叠；在 ViT-Base 与 ViViT-B 上相比最优 VQ 分布式推理框架最高提升 1.89% 精度与 1.27 倍端到端加速，通信开销最高降低 22,624 倍。
Qualcomm CEO Cristiano Amon says leading AI labs want to run at least 100B-parameter models on phones by 2028, as he discusses on-device AI agents, memory price pressures hitting smartphone sales, smart glasses, Modular's open-sourcing, and 6G plans for the 2028 LA Olympics.
作者探讨了一种面向本地受限硬件（<6GB VRAM）的解耦式多智能体架构：用单个中等规模推理模型作为编排者，配合多个1-3M参数的CPU端「纳米模型」处理高频例行任务，仅在低置信度时向上升级，实现零云、空气隔离的边缘优先系统。
The article introduces an edge-based multi-agent orchestration approach using Pi Durable Objects on Cloudflare: each session is mapped to a stateful Durable Object, enabling state persistence, sleep/wake cycles, and parallel sub-agent scheduling — eliminating the traditional heavy backend architecture built on Redis and message queues.
The paper introduces EdgeKernelBench and the Layered Edge Kernel Agent (LEKA) for unified evaluation and layered optimization of LLM-generated kernels across heterogeneous edge backends, finding a large gap between correctness (64.5% Correct@1) and performance (25.5% [email protected]), with LEKA reaching 43.5% [email protected] versus 13.0% for direct generation.
TheAhmadOsman announces a follow-up interview on Local AI, responding to audience interest in their previous discussion on the topic.
论文研究了一种混合架构：云端大模型预先编写程序，将涉及用户私有数据的运行时判断委托给本地小模型（SLM）通过普通函数调用来解释。在 AppWorld 基准上，该方法在所有六种模型与数据分区组合中均提升了纯代码 agent 的成功率（最高 14.9 个百分点），在保持数据本地化的同时缩小了与云端 ReAct 循环的差距。
开发者介绍 NIDO，一个完全在 Android 手机上离线运行的开源 AI 助手，通过确定性路由让小模型（0.5B-7B）只处理真正开放式的任务，并发布了 0.1.29 至 0.1.32 四个 alpha 版本，新增多语言短语路由、重复提醒等功能。
Flox Intelligence, a KTH spin-out and TechCrunch Disrupt Startup Battlefield 200 finalist, is building edge AI devices that detect animals near trains, airports, and roads and emit species-appropriate deterrents to prevent wildlife strikes.