Tag
提出 EC2-Bench，一个评估长时程 LLM 智能体边云协作的基准，揭示在任务性能、资源使用与隐私暴露之间的权衡，并发现更强的边端模型会将协作方式推向更多由边端控制的策略。
Anthropic reportedly confirmed that small 9B AI models may soon be capable enough to self-replicate on commercial IoT devices, with a 9B model achieving inference on 8 of 9 tested hardware types, marking a milestone for local AI.
llama.cpp 项目介绍：一个以 C/C++ 实现的轻量级 LLM/VLM 推理引擎，支持多种硬件后端，可在从服务器到树莓派等各类设备上运行，并提供命令行工具与 OpenAI 兼容的 API 服务。
该推文推荐了一篇非常详尽的 Cloudflare 上手教程，涵盖了 DNS 解析、橙云/灰云、SSL/TLS、邮件路由转发、R2 对象存储、AI 爬虫检测、Workers AI 资源以及用 Agent 接管 Cloudflare 等实用技巧，作者表示跟随练习近一小时收获颇丰。
Deno 宣布整个团队加入 Cloudflare：Deno 运行时将在一年维护期后停止开发，Deno Deploy 六个月后关闭，JSR 将迁移至 Cloudflare 基础设施，未来开发重心转向基于 Cloudflare Workers 和 Durable Objects 的统一平台（celld）。
该论文提出了一种自适应多判别器 Wasserstein GAN（MD-WGAN）框架，将强化学习与博弈论协调结合，用于资源受限的车联网环境，实现与最先进 GAN 相当的预测精度（RMSE 1.029），同时显著降低 CPU 与内存消耗。
Edge0 开源一个月后展示了 35B 模型在 iPhone 飞行模式下以 1.8GB 峰值内存运行的能力，无需云端或服务器，已扩展到耳机、智能眼镜、机器人等设备并拥有超过 8 万用户。
Cascadia 论文展示了在11台 Intel Core Ultra X7 358H AI PC 上分布式常驻推理 975B 总参数（41B 激活参数）的 Inkling MoE 模型，通过自定义 MoE 引擎与 OpenVINO 融合算子实现 60.29 聚合 decode tokens/s 的性能。
该研究评测了轻量级开源语言模型在资源受限的边缘环境中对智能数据模型(SDM)进行分类的性能，涵盖通用、推理专用和代码专用架构，并与TF-IDF及轻量句子编码器基线进行对比，为边缘部署的模型选择和任务设计提供参考。
An EV charging company plans to deploy 100,000 Nvidia GPUs in pods at charging sites across the US, aiming to launch the 'world's first edge inference compute network using idle EV charging capacity' by leveraging underused power at EV stations.
Reposted and commented on @elliothux's open-source open-compute project: it packages the entire Cloudflare Workers platform into a single zero-dependency binary, with KV, D1, R2, DO, Queues, Workflows, and the full ecosystem built in. It uses extremely few resources and can be deployed in one minute.
An interview clip with 0xSigil discussing 'Underdog's Law' — the observation that local AI running fully on consumer devices like a MacBook can already outperform frontier models released by OpenAI and Anthropic just six months prior, without trading off privacy for capability.
Comment on the launch of 'Core,' a personal AI computer, highlighting that its GPU alone is worth $3,200 alongside 64GB DDR5, NVMe storage, and pre-configured software.
Nishant Joshi argues that personal devices on users' desks already hold massive unused compute (≈15 TFLOPS, 46 GB RAM), and that shipping data to the cloud for work whose context is already local is pure waste. He contends servers should only be used for coordination, authority, and sync, though he acknowledges limits around engineering cost, billing trust, and frontier models.
A study investigating post-training quantization (PTQ) on pre-trained AI weather emulators (DLWP and FourCastNet) to accelerate autoregressive inference and enable deployment on edge hardware, finding that quantized configurations retain qualitatively meaningful short-range forecasts.
Cactus Compute releases Whistle, a 16.9 MB open speech recognition model running dependency-free on CPU across mobile, wearables, robotics, automotive and microcontrollers. It transcribes seven languages with word timestamps and speech embeddings, reaching first token in 11 ms, and shares Needle's C++ engine and quantisation.
A tweet argues that local AI setups like DGX Spark and Mac Studio now rival cloud AI on speed, cost-effectiveness, and reliability, offering privacy and freedom from censorship despite being slightly behind frontier cloud models. The quoted reply counters that DGX Sparks are slow, overpriced compared to renting cloud GPUs, and likely to depreciate rapidly.
Cloudflare announced K2 in public beta, a durable, fully serverless event streaming primitive on its Developer Platform that implements a partitioned durable log on top of R2 object storage, enabling scalable event decoupling similar to Kafka but designed for the edge.
AgBench introduces a benchmark suite and open artifacts for evaluating agentic AI on personal devices, comparing local, hybrid, and cloud execution across task success, latency, API cost, and data exposure using over 162 million data points. The study finds no single architecture dominates, with local execution reducing cloud cost and privacy risk but generally lowering task success.
Netlify rebuilt Edge Functions infrastructure on Firecracker MicroVMs (with Unikraft), cutting warm-invocation latency from 25–40ms to ~5–6ms at p50 — roughly 5x faster — while improving security and reliability without changing how developers write functions.