Skip to content

#端侧

0 today

Yesterday · Nov 16Monday

Hacker News front page

How to make almost anything

MIT「How to Make Almost Anything」课程公布 2026 秋季日程,将 AI agents 与 Edge AI 纳入教学:10/01 的 3D 扫描与打印课辅以 AI agents 讲座,10/29 的输入设备课配 Edge AI 幻灯片与代码。课程还覆盖 CAD、嵌入式编程、电子设计与生产、CAM、网络与通信等,12/15 进行期末展示与 open house。

TechCrunch · AI

Indian chip startup Bharat Semi lands $24M seed round as it eyes global AI sensor market

印度半导体初创公司 Bharat Semi 完成 2400 万美元种子轮融资,用于开发 AI 图像传感器。其可编程传感器将图像采集、处理与 AI 能力集成在单芯片上,联合创始人 Vrinda Kapoor 称相比传统方案可降低至少 70% 系统成本。公司已获 General Atomics 和 Kaynes Semicon 的设计订单,首代传感器处于原型阶段,目标 2027 年首次商业量产。

Nov 14Saturday

Hacker News front page

Ggml v1.0.0

ggml 发布 v1.0.0,移除废弃公共 API 和旧版 DIAG_MASK 算子,新增可配置 YaRN 截断与 FP8 支持,并将 RPC 协议升至 v9。

Nov 13Friday

Hacker News front page

Edge0 open-sources Edge0-145B and Edge0-120B for local inference on personal computers

Edge0 released Edge0-145B, Edge0-120B, a unified inference framework and deployment tools, all open source, so coding, reasoning and agent tasks can run on supported personal computers. Internal tests report Edge0-145B at 29GB peak memory, 1,100+ tokens/s prefill and 80+ tokens/s decode. Expert matrix sharing, layered loading and CPU/GPU co-execution cut the runtime load.

Why it matters: Edge0 designed the model architecture and inference engine together, using expert matrix sharing, layered loading and heterogeneous scheduling to cut memory use and data movement for local runs.

Hacker News front page

Foundation Models framework

Apple 的 Foundation Models framework 将原生 Swift API 扩展至云端模型,支持 Apple Foundation Models、Claude 和 Gemini,并提供工具调用、引导生成与图像理解。

TechCrunch · AI

How removing Jensen Huang as Nvidia CEO could hurt the AI industry: five areas

TechCrunch lays out five areas that could be affected if Jensen Huang is removed as Nvidia CEO: open-weight models, agent safety, how people view AI tools, AI PCs and data center financing. The piece argues Huang has been an important advocate for these directions inside the industry.

Why it matters: Huang pushed open-weight models, agent safety and data center financing, and the article asks what loses its main industry advocate if he goes.

Hacker News front page

I'm making LLMs work on small and old devices

smolmaxxing 通过定制内核、内存优化和权重流式读取,让小米 TV Box S(3rd Gen)与 iPad Air(2013)运行最高 16B 模型。

Nov 11Wednesday

Nov 10Tuesday

Hacker News front page

DeepX claims inference up to 100B parameters at less than 5 W

韩国边缘 AI 芯片公司 DeepX 完成第二代 AI 加速器 DX-M2 流片,可在低于 5W 功耗下推理最高 1000 亿参数模型,支持 20B-100B 参数(约 3B-20B 激活参数)模型以 10-30 tokens/s 运行。

Nov 9Monday

Hacker News front page

Ask HN: What are you working on? (November 2026)

Hacker News 本月「Ask HN: What are you working on?」帖中,开发者分享了多个个人项目:Mac 隐私支出追踪器 Meli 可解析信用卡提醒邮件并支持 7 家美国银行及 PayPal、Venmo;开源公交软件 OneBusAway 正筹备推出用普通智能手机替代专有车载定位系统的 Vehicle Positions。

Nov 8Sunday

Hacker News front page

Show HN: Maple – Sub-2-BPW Moe Pruning to Run Gemma 4 26B at 45 TPS on a 6GB GPU

Maple 发布领域专用 MoE 剪枝流水线及 Gemma 4 26B-A4B 压缩模型,将模型从 9.92 GB 缩至 4.75 GB。项目通过前向激活分析每层移除 80 个专家、保留 48 个专家,报告在 RTX 4050 6GB 笔记本 GPU 上实现全部 30 层驻留显存,生成速度由 6.9 提升至 45.4 Tokens/Sec。

Hacker News front page

MeccanicOS

MeccanicOS 是基于 NixOS 构建的一体化 Linux 系统,提供一键制作可启动 USB 的工具,支持 Windows、macOS 和 Linux。

Hacker News front page

Show HN: An offline desktop companion to help maintain a healthy blink rhythm

EyesRhythm 1.5.0 是一款完全本地运行的桌面眨眼监测工具,用摄像头统计眨眼频率与规律性并折算成 0-100 的 Ocular Rhythm Index。它提供实时眨眼曲线、每 30° 分片的 Poise 头部朝向表盘和两分钟 Practice 引导,数据不上传、无需账号。支持 macOS 与 Windows,免费下载,安装包分别为 93 MB 和 84 MB。

Hacker News front page

What I run on my $10 USD OpenStick homelab

一台 10 美元(含运费)的 OpenStick 被作者用作家庭服务器,硬件为 512 MB 内存、4 核 1 GHz ARM Cortex-A53、4 GB eMMC 和 4G 模组,运行 Debian 与 Podman。

Nov 7Saturday

Nov 6Friday

TechCrunch · AI

Underdog shows off a 32MB speech model in new demo

Underdog 推出基于新语音生成模型 Woof 的浏览器游戏演示,该模型大小仅为 32MB,拥有 31 million 参数并开放下载。用户可在游戏中的打字机输入文字并听到朗读,也可旋转调谐器生成不同音乐;公司称模型可在大多数设备上无需联网运行。几周前,Underdog 还推出了面向端侧转录和听写的 167 MB 语音转文字模型 Bark。

Nov 5Thursday

TechCrunch · AI

I tried Apple’s Siri Recaps. They’re surprisingly useful — and a little unsettling

Apple 的 Siri Recaps 在实测中记录了修车和租车的关键口头信息,也混入家庭闲聊与不必要的细节。该功能随 iOS 27.3 和 watchOS 27.3 开发者测试版进入测试,可为 Apple Watch Series 12 用户汇总全天音频;Apple 尽可能在端侧处理,不识别说话者、不保存原始音频,并隐去私人信息。

Nov 4Wednesday

Hacker News front page

Preparing for the Flipper One Kickstarter and opening pre-orders

Flipper One 开启预售,这是一款基于 Rockchip RK3576 八核 ARM SoC 的开源 Linux 网络多功能工具,搭载 Raspberry Pi RP2350 MCU 双处理器架构,内置 Wi-Fi 6E 与千兆以太网,支持 M.2 模块扩展 5G、SDR 与 AI 加速器。

Simon Willison

Qwen 3.8 27B on my laptop can interpret model weights, too

Simon Willison 在笔记本上运行 Qwen 3.8 27B,仅凭原始权重、无工具调用,成功识别三个游戏网络的用途。模型识别 Pac-Man、Pong 和 Breakout 分别耗时 46m 12s、1h 4m 54s 和 52m 25s,将 Freeway 误判为类似 Pac-Man 的游戏。

The Verge · AI

Apple bets on AI across its own and third-party tech to grow revenue

Apple turns AI use into revenue through hardware sales, a cut of third-party AI app subscriptions, and planned paid tiers of its own AI services. Last quarter's capital expenditure was $2.2 billion, against $77.3 billion for Amazon and $51.4 billion for Alphabet; those figures cover total capex, not AI alone.

Why it matters: Apple makes money from AI through hardware, app subscription cuts and paid tiers of its own services, while spending far less on capex than peers building infrastructure.

Nov 3Tuesday

Hacker News front page

Show HN: Subvocal – Talk to your computer without making noise

Subvocal 展示可将无声口部动作解码为文本的智能眼镜,目前已接受预订。团队在讨论中解释,系统通过 RF 感知口腔内部运动,经手机或笔记本电脑上的配套应用使用云端解码,端到端延迟通常低于 500ms。初期计划通过连接设备上的按键触发输入,自动检测无声说话活动的功能仍在研发,预计在发货几个月后通过软件更新提供。

Google Blog · Gemini

Google Recaps October 2026 AI Updates, Covering Gemini 4 and Genie 4

Google's October 2026 AI recap covers the global rollout of Gemini 4 Argon, the announcement of Gemini 4 Carbon, and the world model Genie 4. Argon goes out worldwide while Carbon stays in limited testing.

Why it matters: Google is running Argon's global release and Carbon's limited test on separate tracks, so the two frontier models differ in access and safety-testing stage.

Nov 1Sunday

Hacker News front page

Show HN: A language model that lives in a boot sector

有人用 512 字节手写汇编 MBR 加载 11.5 KB int8 模型,在 386 的 16 位实模式下运行,无需操作系统和 C 运行时。模型仅有 8 字符上下文和一层隐藏层,启动后会"梦见"一段并不存在的 Linux 启动日志,包含时间戳、PCI 地址和 ACPI 事件。权重、字符串和词表存放在后续 23 个扇区,Python 训练脚本与汇编推理逐字符对齐。

Computing Life · Share · Yage

从 Liquid AI LFM3 看端侧小模型的同源分工策略与工程取舍

Liquid AI 的 LFM3 家族通过共享基底、按任务分支和削减全局注意力,让端侧小模型专注工具执行与结构化决策。10 月 27 日开源的六个模型共享 LFM3-1.7B-Base,分为四个生成模型和两个零输出 token 的决策模型;LFM3-1.7B 的闭卷知识准确率为 7.07%,未知拒答率为 92.93%。

Oct 31Saturday

Hacker News front page

Digital minimalism

数字极简主义主张有意识地使用数字技术,博客作者通过社交媒体、蜂窝网络和网页使用实验探索减少数字依赖。作者改用 Firefly 访问社交平台,并进行持续 100 天、保留 wifi 的手机飞行模式实验,认为体验总体成功,但遇到短信验证和出行不便。

The Verge · AI

Sam Altman says RAM belongs in data centers, not personal computers

On the Cheeky Pint podcast, OpenAI CEO Sam Altman said tight RAM supply should go first to raising cloud utilization. Spreading limited memory across personal computers is not ideal, he said.

Why it matters: Altman uses low memory utilization on personal computers to argue for concentrating RAM in the cloud, as DRAM supply stays constrained.

Oct 30Friday

Hacker News front page

A tiny world model running games on a Raspberry Pi Pico 2 W

tinyworldformer 在 Raspberry Pi Pico 2 W 上用世界模型生成游戏帧,再由另一小型网络选择动作来玩游戏。世界模型采用共享权重的循环 Transformer 架构与 int8 量化,支持 Breakout、Doom-lite、Flappy Bird、Pong 和 Pac-Man lite,也支持在设备上从零训练智能体。

Oct 29Thursday

Google Blog · Gemini

Google and XREAL launch XREAL AURA, a wired Android XR glasses for work, video and games

Google and XREAL launched XREAL AURA, a wired XR headset running Android XR. Gemini can read the real-world view and on-screen content and help with it. The glasses weigh under 95 grams with a 70-degree field of view, connect by cable to a portable compute module, support multi-window apps and USB-C laptop screen mirroring, and Gemini can highlight objects in view.

Why it matters: AURA puts Gemini into the real-world view and laptop windows, so users can ask about what is in front of them while working across apps.