OpenClaw🦞@openclaw · 3小时前29

Thank you, @colinsolvely, for helping move the needed iOS and Android UI improvements forward with this post. We want to meet this feedback in the same spirit. Please share what is not working here or in Discord so the team can listen, respond, and turn that feedback into improvements.

译OpenClaw 项目官方发推感谢 @colinsolvely（Colin）推动 iOS 和 Android UI 改进，并鼓励用户继续在 X 或 Discord 反馈问题。Colin 在引用推文中自我介绍，称自己是负责 OpenClaw iOS UI 的志愿者维护者，时间有限但热衷移动端 AI 智能体，坦承 UI 有问题并接受批评，邀请大家在 Discord 对话，共同改进。双方强调社区共建。

Berryxia.AI@berryxia · 8小时前48

卧槽,手机就可以完成3D建模了！ GenRecon提出了一种把生成式3D先验和多视角重建结合起来的新方法。 它不再单纯依赖传统SfM/MVS或NeRF-style优化，而是把场景切成有重叠的chunk，用强生成模型（比如Trellis.2）做条件生成来重建每个chunk，再拼起来。 核心创新是用投影式的conditioning机制，把多视角图像特征直接提升到和生成模型对齐的3D空间里。 最终输出是高质量、可编辑的PBR mesh，在室内场景重建上据称比当前SOTA高出16%的保真度和完整度。 这其实代表了当前3D重建的一个趋势：不再只靠几何约束，是越来越多地借用生成模型的先验来补全缺失信息、提升细节。

译GenRecon将生成式3D先验与多视角重建结合，把场景切分成重叠chunk，用Trellis.2等生成模型条件重建各chunk并拼接。核心创新是投影式conditioning，将多视角图像特征提升到3D空间。输出可编辑PBR mesh，室内重建保真度和完整度比SOTA高16%。

Chubby♨️@kimmonismus · 9小时前55

Huge:Anthropic is reportedly in early talks with Samsung to manufacture its own custom AI chip. The Information says the Claude maker is still figuring out what the processor should do, how powerful it should be, and how it would fit into servers or clusters. Anthropic already uses AWS Trainium, Google TPUs, and Nvidia GPUs, and says those will remain central. A custom chip would give it another lever as model deployment costs, data-center capacity, memory supply, and power all become strategic constraints. Samsung would get a high-profile AI foundry customer for its 2-nanometer process and advanced packaging push.

译The Information报道，Anthropic正与三星早期洽谈制造自研AI芯片，目前仍在规划芯片功能、算力及服务器/集群适配方案。Anthropic现有AWS Trainium、Google TPU、Nvidia GPU仍为核心，自研芯片将作为应对模型部署成本、数据中心容量、内存供应、电力等战略约束的新杠杆。三星借此获得高端AI代工客户，推动其2纳米制程和先进封装业务。

郭明錤｜Ming-Chi Kuo@mingchikuo · 18小时前55

This structural shift suggests that Amazon has been preparing for the long-term expansion of AI compute, and that it does not see AI compute oversupply as a real issue.

译郭明錤调研显示，Amazon消费电子（Kindle、Fire TV、Echo等）将首次改变20年来外购处理器策略，转向COT自研模式，由Alchip独家负责后端设计与测试。背景是AI算力快速扩张导致截至2026年Q1的12个月自由现金流下降95%至约12亿美元，Amazon需优化非AI业务成本以维持AI投资。新策略预计2027年启动，完成后年自研处理器出货约4000万。主推文认为，Amazon正为AI算力长期扩张做准备，并不认为AI算力过剩是真实问题。

郭明錤｜Ming-Chi Kuo@mingchikuo · 19小时前49

My latest industry checks indicate that Amazon’s processor procurement strategy for its own consumer electronics is set to undergo its first major shift in 20 years: moving away from externally sourced processors and adopting a COT (customer-owned tooling) model, with Alchip as the exclusive provider of back-end design and testing for its self-developed chips. Key points: 1. Amid the rapid expansion of AI compute, Amazon’s free cash flow for the 12 months ended 1Q26 fell 95% year over year to about US$1.2 billion. To maintain financial flexibility and keep funding its AI investment cycle, Amazon is also streamlining its organization and improving the cost structure of its non-AI businesses. 2. Amazon's own-brand consumer electronics, including Kindle, Fire TV, Echo, Alexa-enabled products, Blink, and Ring, currently use externally sourced processors. To optimize its cost structure, Amazon plans to gradually shift away from external sourcing and adopt a COT model similar to the one used for AI chip Trainium, taking more control of processor development in-house. 3. For this transition, Amazon has selected Alchip as its exclusive partner for back-end design and testing. Alchip is expected to receive non-recurring engineering, or NRE, fees for each design project, while also benefiting from processor shipments. 4. The strategy is expected to begin in 2027. Once the transition is complete, annual shipments of Amazon’s in-house processors are estimated at around 40 million units. This shift should help improve the cost structure of Amazon’s own devices and become a meaningful growth driver for Alchip.

译郭明錤产业调查显示，Amazon拟对其消费电子处理器采购进行20年来首次重大调整。因AI计算扩张，截至1Q26的12个月自由现金流同比降95%至约12亿美元，为优化非AI业务成本，Amazon将逐步放弃Kindle、Fire TV、Echo等自营设备的外部处理器，采用类似Trainium的COT模式，将处理器开发内包。已选定Alchip为独家后端设计与测试合作伙伴，收取NRE费用并受益于芯片出货。策略预计2027年启动，完成后年出货约4000万颗，有望改善设备成本并成为Alchip增长动力。

Berryxia.AI@berryxia · 21小时前42

Space 𝕏 的版图和老马的叙事不是盖的。 可能大家陆续看到各版块𝕏的产品出现要。

译据《华尔街日报》报道，SpaceX正在开发一款AI硬件产品，比iPhone更薄。主推文感叹：Space 𝕏的版图和马斯克的叙事确实厉害。

Berryxia.AI@berryxia · 1天前63

Wow ,这个项目还是有点东西啊！ 真正不离开你电脑的AI 工作台 ，真的可以试试~ 目前6k GitHub Stars，15.5万次下载，MIT开源！ 来自Osaurus @OsaurusAI App做了一个挺有意思的设计。 它不是浏览器套壳调用云端，直接是真正用Swift在本地跑图像生成。 用户要生成图片时，它会自动把当前聊天模型卸载，加载本地图像模型（支持任意MLX模型），在GPU上生成，保存真实文件后再切回聊天模型。 整个过程完全本地、无需联网，而且切换很丝滑。 作者强调“Most Mac AI apps are a browser calling the cloud”，而Osaurus想做的是真正的本地原生体验。 目前免费开源，适合想在Mac上本地跑多模型切换的用户。 地址见评论区👇🏻，我做了个视频介绍可以看看！ 记得一键三连啊！

译Osaurus AI App 推出 Mac 原生 AI 工作台，不是浏览器套壳，而是用 Swift 在本地运行图像生成。用户生成图片时自动卸载当前聊天模型，加载本地图像模型（支持任意 MLX 模型），在 GPU 上生成并保存真实文件后再切回聊天模型，全程无需联网。目前 6k GitHub Stars、15.5 万次下载，MIT 开源，适合在 Mac 上本地跑多模型切换。

OpenBMB@OpenBMB · 1天前35

Really well-written and practical. 👍We appreciate how you broke down the entire local pipeline and highlighted why MiniCPM-V 4.6 is such a strong choice for edge deployment — the combination of tiny size, strong multimodal capability, and excellent efficiency on Apple Silicon is exactly what makes these models exciting.🤗

译真的写得很好且实用。👍 我们很欣赏你如何拆解整个本地管道，并突出说明了为什么 MiniCPM-V 4.6 是边缘部署的绝佳选择——其小巧的尺寸、强大的多模态能力以及在 Apple Silicon 上出色的效率正是这些模型令人兴奋之处。🤗

小互@xiaohu · 1天前36

Vibe Coding 大杀器来了，有点意思 告别高声自言自语的尴尬，小声默念就能自动识别你的声音并进行语音输入 一款智能戒指：轻声低语即可语音书写内容 而且轻轻触摸戒指即可进行编辑 还可以通过手势（如轻弹手指）在不同的应用程序、设备和 AI 之间快速切换与联动 单次充电可使用 16 小时... 原生支持 iPhone、Mac、Vision Pro 等苹果设备

译一款智能戒指可实现小声默念语音输入，告别高声自语；轻触戒指即可编辑，通过弹指等手势在不同应用、设备和AI间快速切换联动。单次充电续航16小时，原生支持iPhone、Mac、Vision Pro等苹果设备，成为Vibe Coding场景下的便捷工具。

Rohan Paul@rohanpaul_ai · 2天前56

love it. Claude desktop app comes to Ubuntu/Linux. I'm a hardcore Ubuntu user, and till now could only use Claude through the browser and terminal. Claude Mythos proly built and released this in about an hour 😂

译太爱了。Claude 桌面应用现已登陆 Ubuntu/Linux。 我是一名硬核 Ubuntu 用户，之前只能通过浏览器和终端使用 Claude。 Claude Mythos 大概在一小时内就构建并发布了这个版本 😂

ClaudeDevs@ClaudeDevs · 2天前67

Claude Desktop is now available on Linux (Ubuntu and Debian) in beta. Alongside the browser and terminal, you now get a first-class desktop experience with Claude Code, Claude Cowork, and chat on all paid plans.

译Claude Desktop 现已在 Linux（Ubuntu 和 Debian）上推出测试版。 除了浏览器和终端，你现在可以在所有付费计划中获得一流的桌面体验，包括 Claude Code、Claude Cowork 和聊天。

Rohan Paul@rohanpaul_ai · 2天前40

Satya Nadella just talked about a new AI Dev Box (RTX Spark Dev Box) with 20 CPU cores, 128GB unified memory, and petaflop of local compute for trillion-parameter-scale model runs. ---- From "Stanford Online" YouTube channel, (link in comment)

译Satya Nadella 刚刚谈到了一款新的 AI Dev Box（RTX Spark Dev Box），配备 20 个 CPU 核心、128GB 统一内存和千万亿次本地计算能力，用于运行万亿参数规模的模型。

OpenBMB@OpenBMB · 2天前54

Thanks for the shoutout! 🤗 @HuggingModels Exceptional fine-grained OCR, complex image reasoning, and multi-turn interaction in a highly compact footprint. Fully open-sourced with out-of-the-box support for SGLang/vLLM/llama.cpp/Ollama, multi-platform mobile deployment, and low-barrier fine-tuning on consumer GPUs. https://huggingface.co/openbmb/MiniCPM-V-4.6

译面壁智能回应HuggingModels的推荐，介绍了MiniCPM-V-4.6多模态模型。该模型具备精细OCR、复杂图像推理和多轮交互能力，尺寸紧凑，完全开源。它开箱支持SGLang、vLLM、llama.cpp、Ollama等推理框架，可部署于多平台移动端，并支持在消费级GPU上进行低门槛微调。引用推文强调这是一款能同时理解文本和视觉的轻量级AI模型，适合设备端使用，无需依赖云端。

歸藏(guizang.ai)@op7418 · 2天前61

Cursor 的 iOS 版上线了，完成度很高啊 锁屏的时候，还会展示当前的这个进度，然后也完成以后，会把界面的视频和图片给你发过来，让你审核 当前，针对付费用户的 beta 测试已经开放

OpenClaw🦞@openclaw · 3天前71

OpenClaw is now on iOS + Android 🦞 📱 Native mobile apps, finally 💬 Agents in your pocket 🔔 Channels, tasks, replies on the go Run agents from wherever your thumbs are. iOS: https://apps.apple.com/us/app/openclaw-ai-that-does-things/id6780396132 Android: https://play.google.com/store/apps/details?id=ai.openclaw.app

译OpenClaw 现已登陆 iOS 和 Android 🦞 📱 原生移动应用，终于来了 💬 智能体装进口袋 🔔 频道、任务、回复，随时处理 用你的大拇指，在任何地方运行智能体。 iOS: https://apps.apple.com/us/app/openclaw-ai-that-does-things/id6780396132 Android: https://play.google.com/store/apps/details?id=ai.openclaw.app

AYi@AYi_AInotes · 3天前68

Cursor iOS端上线了， 中国的iOS用户今晚可以甜甜入睡了😴

译Cursor for iOS 正式发布，支持启动云端智能体，或从 App 远程控制电脑上的智能体。Composer 2.5 即日起至 7 月 5 日享 75% 折扣。 中国的iOS用户今晚可以甜甜入睡了😴

MiniMax (official)@MiniMax_AI · 3天前39

This is a glimpse of where local AI is heading and we are glad to be part of it. Really impressive work by all the teams involved @Gradient_HQ, @tryParallax, and @GA_agent_ai

译MiniMax官方转发了Gradient、Parallax和GenericAgent团队的演示结果。他们在本地运行了MiniMax M3（428B参数模型），通过Parallax工具部署在3台Mac上，再由GenericAgent驱动一个约3000行代码的自主智能体，完成了创建5只股票投资组合并写入磁盘的任务。整个过程完全在本地进行，无云端调用、无API费用，数据未离开机器。MiniMax表示这是本地AI未来发展的一个缩影。

郭明錤｜Ming-Chi Kuo@mingchikuo · 6天前58

iOS 27 將進一步深化與 Apple Intelligence 的系統級整合。我的最新產業調查顯示，為確保在 AI 負載下系統仍能順暢運作，1H27 新款配備 A20 處理器的較低階 iPhone 之 DRAM 規格將升級至 9GB（1.5GB × 6-die），高於目前 A19 機型的 8GB（2GB × 4-die）。2H26 配備 A20 Pro 處理器的三款高階機型（折疊機與兩款 18 Pro 機型）則維持 12GB（1.5GB x 8-die）。

译郭明錤最新产业调查指出，iOS 27将深化与Apple Intelligence的系统级整合。为确保AI负载下系统顺畅，1H27新款搭载A20处理器的较低阶iPhone DRAM规格升级至9GB（1.5GB × 6-die），高于当前A19机型的8GB（2GB × 4-die）。2H26配备A20 Pro的三款高阶机型（折叠机与两款18 Pro）则维持12GB（1.5GB × 8-die）。

郭明錤｜Ming-Chi Kuo@mingchikuo · 6天前48

iOS 27 will bring tighter system-level integration with Apple Intelligence. My latest industry checks suggest Apple's lower-end 1H27 iPhones, powered by the A20 chip, will move to 9GB DRAM (1.5GB × 6 dies), up from 8GB (2GB × 4 dies) in the current A19 models, to keep the system running smoothly under AI workloads. The three new high-end 2H26 models powered by the A20 Pro chip (the foldable and two 18 Pro models) will remain unchanged at 12GB (1.5GB × 8 dies).

译iOS 27 将带来与 Apple Intelligence 更紧密的系统级集成。我最新行业调查显示，苹果 2027 上半年搭载 A20 芯片的低端 iPhone 将升级至 9GB DRAM（1.5GB × 6 颗粒），高于当前 A19 机型的 8GB（2GB × 4 颗粒），以确保系统在 AI 负载下流畅运行。三款搭载 A20 Pro 芯片的 2026 下半年高端新机型（折叠屏及两款 18 Pro 机型）则维持 12GB（1.5GB × 8 颗粒）不变。

OpenBMB@OpenBMB · 6天前41

We’re excited to see MiniCPM5-1B being used in real NAS-based local AI systems.🥳🥳 A developer in our community built a full-stack setup combining on-device LLM inference with NAS and Agent capabilities: ⚡ Lightweight local deployment MiniCPM5-1B runs on a QNAP-Qu605-N150-16G NAS, consuming <2GB of memory. It is deployed via Ollama and integrated into Cherry Studio as a local LLM provider. 🧩NAS + Agent integration via MCP With NAS MCP, system capabilities like file management , shared folders , and semantic search are exposed to external agents. This enables Coding Agents / WorkBuddy-style workflows to securely access and retrieve local data within permission boundaries. 📚Local knowledge base+ RAG pipeline Using Qsirch indexing, NAS files can be turned into a structured local knowledge base. MiniCPM5-1B handles retrieval-based reasoning, enabling summarization, Q&A, and extended reasoning fully on-device. This is a great case of how efficient small models are evolving beyond local inference into real system-level intelligence. From NAS storage → Agent operations → knowledge reasoning Everything works together in one loop! 📖 Original post: https://mp.weixin.qq.com/s/iBeHfOrwulYsEm2hhhv7vw

译面壁智能社区开发者将MiniCPM5-1B部署于QNAP NAS（型号Qu605-N150-16G），内存占用低于2GB，通过Ollama集成至Cherry Studio作为本地LLM。借助NAS MCP协议，将文件管理、共享文件夹、语义搜索等系统能力暴露给外部Agent，实现安全本地数据访问。同时利用Qsirch索引将NAS文件构建为结构化知识库，由MiniCPM5-1B在设备端执行检索增强推理，完成摘要、问答等任务。展示轻量小模型从本地推理向系统级智能体+RAG组合演进的实践。

Google AI Developers@googleaidevs · 7天前55

When we hit 2 million apps, we have to build the real statue, right?

译AI Studio 原生 Android 应用构建功能于 2026 年 5 月宣布，目前已有超过 100 万款 Android 应用在 AI Studio 中创建。主推文调侃：等达到 200 万时，是不是该建一座真雕像？

Chubby♨️@kimmonismus · 7天前60

Apple is making one of its biggest Mac silicon strategy shifts yet. According to Bloomberg, Apple plans to launch a base M6 chip as early as this year, but skip the usual M6 Pro and M6 Max variants entirely. Instead, the company is reportedly moving its next high-end Mac chips directly to the M7 generation in 2027. The reason: Apple wants to fast-track more advanced on-device AI and graphics capabilities. The M6 is expected to bring higher memory bandwidth, an upgraded Neural Engine, improved CPU cores, better video encoding and decoding, and a redesigned GPU with up to 12 graphics cores. Memory bandwidth is becoming one of the key specs for AI workloads. The M6 is reportedly targeting around 200 GB/s, up from about 153 GB/s on M5. The base M7 could push that to around 240 GB/s. Apple is also still planning an M5 Ultra for a new Mac Studio, with around 36 CPU cores, 80 GPU cores, and support tested for up to 768 GB of memory. Apple seems to be reorganizing its Mac silicon strategy around a very clear assumption: on-device AI will require much more memory bandwidth, stronger local inference, and better graphics performance than the current Mac roadmap was originally built for. really excited for the m7 chips. My assumption: primarily because the new CEO John Ternus was the significant reason for Apple's shift towards its own M-chips and is now placing an even stronger focus on them.

译据彭博社报道，苹果计划最早今年推出基础M6芯片，但跳过M6 Pro和M6 Max，直接于2027年进入M7代。主因是加速设备端AI和图形能力。M6将带来更高内存带宽（约200 GB/s，M5约153 GB/s）、升级神经引擎、改进CPU核心、增强视频编解码及全新GPU（最多12核心）。M7内存带宽或达约240 GB/s。同时计划推出M5 Ultra用于新Mac Studio，配置约36 CPU、80 GPU核心，支持最高768 GB内存。苹果认为设备端AI需要更高带宽、更强本地推理和图形性能。

Logan Kilpatrick@OfficialLoganK · 7天前61

Gemma 4... intelligence for everyone on device!

译Gemma 4... 为每个人带来设备端智能！

Berryxia.AI@berryxia · 7天前60

这下让真的可以让很多人都闭嘴了！ Unsloth把GLM-5.2压缩到1-bit后。 本地跑起来居然还能和Claude Opus、GPT-5.5正面比创意输出。 他们用Mac Studio M3 Ultra 256GB RAM跑1-bit版本，速度还能到21 tok/s左右。 在同一个prompt下生成的HTML/设计效果，看起来甚至比闭源模型更丰富、更“有想法”。 这已经不是简单的量化了，而是把一个原本需要海量显存的超大模型，硬生生塞进了消费级硬件还能打。 GLM-5.2本身就以创意和长上下文见长，现在连极致量化后都还能保持较强的表现，确实有点超出预期。 这也再次验证了一个趋势：开源模型在极端优化后，正在快速缩小和闭源前沿模型在实际可用性上的差距，尤其是在本地部署和特定任务上。 大内存的本子这下真的太香了，Qwen 3.7 这些模型又该迭代版了。

译Unsloth 将 GLM-5.2 压缩为 1-bit GGUF 量化版本，在 Mac Studio M3 Ultra（256GB RAM）上以约 21.6 tok/s 本地运行。与 Claude 4.8 Opus、GPT-5.5 使用相同提示进行创意输出（HTML/设计效果）对比，1-bit 版本表现不逊色，甚至更丰富、“更有想法”。GLM-5.2 本身以创意和长上下文见长，极端量化后仍保持较强表现，验证了开源模型在极端优化后正快速缩小与闭源前沿模型在实际可用性上的差距，尤其适合本地部署。

OpenBMB@OpenBMB · 6月24日26

Nice work 👍👍@TechlatestNet A local photo assistant on Telegram with MiniCPM-V 4.6 and no cloud OCR is exactly what privacy-focused users need.

译Nice work 👍👍@TechlatestNet Telegram 上的本地照片助手，搭载 MiniCPM-V 4.6，无需云端 OCR，正是注重隐私的用户所需。

Chubby♨️@kimmonismus · 6月24日63

Vida is doing something I don't see often: an agent that runs locally, holds your context from one session to the next, and tries to draft the work before you ask. They're building 100 use cases in public and say they won't really commercialize until at least 20 of them reliably deliver. I like that order of operations.

译主动式 AI 智能体 Vida 今日正式推出。它可完全在本地运行，自动跨会话保持上下文，并在用户提出请求前主动起草工作成果。团队以 #BuildInPublic 方式公开构建 100 个典型用例，首批 5 个已上线。创始人表示，在至少 20 个用例达到可靠交付前不会商业化。Vida 定位于解决 Slack 消息泛滥、网页标签切换、手动总结文档会议等现代工作流痛点，强调“未来 AI 不是更好的对话，而是更好的结果”。免费体验地址：vida.app。

OpenBMB@OpenBMB · 6月24日65

🥳Thanks for sharing this, @MLBoy_DaisukeMajima 🚀 MiniCPM-V 4.6 running at this speed on-device is really impressive — especially under 2B params on Apple Core AI. Great work pushing efficient multimodal AI forward.🫡

译🥳感谢分享，@MLBoy_DaisukeMajima 🚀 MiniCPM-V 4.6 在设备上以这样的速度运行，实在令人印象深刻——尤其是在 Apple Core AI 上以不到 2B 参数跑出。 干得漂亮，推动高效多模态 AI 向前发展。🫡

OpenBMB@OpenBMB · 6月24日63

Big thanks to @JackdeS11 for bringing VoxCPM-0.5B fully on‑device to iPhone! 🎉❤️ The entire stack (MiniCPM4 + LocDiT flow‑matching + AudioVAE) runs on Neural Engine and GPU, with no network required. Great work! 👍👍

译面壁智能（OpenBMB）的扩散式 TTS 模型 VoxCPM-0.5B 已通过 Apple Core AI 完全部署至 iPhone 端侧，无需联网。该模型整合了 MiniCPM4 语言模型、LocDiT flow-matching 和 AudioVAE，每一层均运行于 Neural Engine 和 GPU 上。模型权重和部署代码已开源至 HuggingFace 与 GitHub。

小互@xiaohu · 6月24日68

字节跳动旗下AI聊天应用豆包推出专业版 正式开启付费 包月68元起，高级套餐500元/月 豆包专业版专注应用开发、数据分析等专业工作场景 豆包办公任务模式支持操作本地电脑、使用浏览器、调用Skills技能和定时任务等能力， 内置Office办公套件，并支持专业图片、视频设计，以及应用和网站生成。 专业版由最新的豆包2.1Pro模型驱动，该模型能力和opus 4.6相当，在该模式下，豆包可以理解工作目标，自主拆解任务，并调用本地电脑、文档、表格、网页等工具持续执行任务，从“回答问题”升级为“生产力工具”。 三级阶梯定价（连续包月）：套餐越贵、额度越大。 标准套餐：68元/月，额度为免费版的5倍以上 加强套餐：200元/月，额度为标准套餐的4倍 高级套餐：500元/月，额度为标准套餐的10倍 学生特惠：即将面向在校大学生推出6个月活动，认证后享更多免费额度，标准套餐可享 38元/月 专属价。

译豆包推出专业版，开启付费。三级阶梯定价：标准68元/月、加强200元/月、高级500元/月，额度依次递增。专注应用开发、数据分析等场景，支持办公任务模式，可操作本地电脑、调用Skills、内置Office套件。由豆包2.1Pro模型驱动。学生认证后标准套餐38元/月，活动期6个月。

Rohan Paul@rohanpaul_ai · 6月24日60

Atomic Chat just made Cline run coding agents on local AI models. @cline is the agent layer: it can read files, edit code, run terminal commands, inspect errors, and ask approval before actions. Atomic Chat is the model layer: it runs open-weight models offline locally, supports 1000+ models, and exposes a local API that other tools can call. A quieter shift in many areas: powerful agents moving back onto private machines.

译Atomic Chat 集成 Cline，使代码智能体能在本地 AI 模型上运行。Cline 作为智能体层，可读文件、编辑代码、运行终端命令、检查错误并请求批准，已被 800 万+开发者信任。Atomic Chat 作为模型层，支持 1000+ 模型离线本地运行，并暴露本地 API 供其他工具调用。整个过程私有、免费、开源，体现了强大智能体回归本地私有设备的趋势。

SemiAnalysis@SemiAnalysis_ · 6月24日52

Alibaba's core chip entity behind their PPUs, T-Head, just filed a business registration change lifting registered capital from RMB 300mn to RMB 1 Bn, a >3x increase. The same filing also moved its shareholder from DAMO Academy to PingtouGe (Shanghai) Electronics. Capital topped up and equity structure cleaned up. This comes roughly 6 months after January's market chatter around a T-Head IPO. [Image source - Xiaohongshu 大厂叭叭]

译阿里巴巴核心芯片实体平头哥（T-Head）刚刚提交了工商变更登记，注册资本从3亿元人民币增至10亿元人民币，增幅超3倍。同一份文件还将其股东从达摩院变更为平头哥（上海）电子。资本金到位，股权结构厘清。此事距1月份市场传言平头哥IPO约6个月。 [图片来源 - 小红书 大厂叭叭]

Hao AI Lab@haoailab · 6月24日73

(1/5) 5 seconds of video. 1.8s seconds of generation. One NVIDIA GeForce RTX 5090 on FastVideo. 🤯🚀 - FastWan-QAD, a new family of video generation models - Trained with FastVideo's Quantization-Aware Distillation (QAD) recipe. - Powered by FastVideo, we push a single NVIDIA GeForce RTX 5090 to its absolute limit: generating a 5-second 480P video in 1.8s end-to-end! 📜 Blog: https://haoailab.com/blogs/fastwan-qad/ 💻 Code: https://github.com/hao-ai-lab/FastVideo 💽 Model: https://huggingface.co/FastVideo/FastWan-QAD-1.3B

译Sky Computing Lab 发布 FastWan-QAD 视频生成模型系列，基于 FastVideo 的量化感知蒸馏（QAD）方案训练。在单张 NVIDIA GeForce RTX 5090 上，端到端生成一段 5 秒 480P 视频仅需 1.8 秒。模型、代码及博客已开源。

Rohan Paul@rohanpaul_ai · 6月24日67

The biggest weakness of AI agents: every useful run usually dies with the session. Imagine being able to turn an AI workflow into a desktop application that runs without consuming any tokens when restarted. i.e. non-technical users can create complete, deployable applications through natural language alone No programming background needed. Just open the app, describe what you need, done. No vibe-coding, instead it’s outcome-coding. Newly launched KroWork is turning an AI chat into a small desktop app you can reuse. Instead of getting a one-time answer, you get a workflow that can run again without rebuilding it. So with KroWork, you describe a task once, the AI agent builds the workflow, and then you can save it as software you actually keep. The conversation becomes something persistent, local, and reusable. Besides, it runs purely locally — your data never leaves your device. I want to monitor the market in real time and access stock information, but tools like Yahoo Finance, Seeking Alpha, and SEC filings are scattered, and paid tools with similar features are too expensive. KroWork can directly generate a reusable application to help me monitor the market for free. With just one command, installation and free deployment are complete.

译KroWork 新发布，解决 AI 智能体每次运行后会话即失效的痛点。用户用自然语言描述任务，AI 自动构建工作流并保存为可重用的桌面应用，无需编程背景。应用完全本地运行，数据不出设备。相比一次性回答，它生成持久、可复用的软件——例如一句指令即可生成免费市场监控应用，替代分散的付费工具。KroWork 将“氛围编码（vibe-coding）”转变为“结果编码（outcome-coding）”。

向阳乔木@vista8 · 6月23日48

大聪明（赛博禅心）一直跟我说，做自媒体要研究新智元。 喷归喷，但人家起标题真的是一绝。 百度在一直OCR方面的技术很强，昨天开源 Unlimited OCR更强，3B参数500M激活，这么小的模型竟然有这么好的效果，出乎意料。 先不管作者是不是从DeepSeek出走的，文章提到的参考滑动注意力窗口技术值得关注。 技术解读见第二条，后面有项目开源地址。

译百度开源 Unlimited OCR 模型，仅3B参数、500M激活参数，在小参数量下实现极佳效果。推文提到其参考了滑动注意力窗口技术，并附带了技术解读和开源地址。

karminski-牙医@karminski3 · 6月22日54

想买Mac运行大模型? 这是劝退贴 其实估算方法很简单, 现在买 MacStudio 哪怕运行 Qwen3.6-27B 4bit 量化版本, 然后开 DFlash 使用Qwen的内置投机解码, 也就飙到 65token/s. 而现在普遍大模型都能跑到 40 token/s. 如果专门买 MacStudio M3 Ultra 96G 运行大模型, 如果把设备售价 (32999) 换算成使用API, 以 GLM-5.2 为例, 每百万token 28块, 一台 MacStudio 的价格大概能买到 32999/28 = 1178M token. 而为了输出这些token, 买到的 MacStudio 运行 Qwen3.6-27B 要持续运行 209天. 也就是说回本周期至少是200天不间断运行. 然后运行模型才是纯赚. 这还是没算电费和不直接买API而是买套餐的情况.而且, 最重要的是这还是在运行一个只有27B的小模型. 如果真的买512G的 MacStudio (108749, 而且好像已经断货了), 然后运行量化版本的 GLM-5.2, 速度就会跌到只有 17 token/s, 回本周期大概在 7 年左右... 对于现在1.5个月模型就发新版本的情况下, 普通用户自用是绝对不划算的. 所以大部分用户买 coding plan 会更划算, 如果像我一样要测新模型, 直接租卡也会比直接买划算很多. 当然, 如果你本身就有Mac或者显卡, 那么空闲的时候(比如睡觉的时候)让它跑大模型运行任务, 反而是划算的. #本地大模型 #mac #qwen36 #glm52

译买MacStudio运行大模型性价比不高。以M3 Ultra 96G（32999元）为例，运行Qwen3.6-27B 4bit量化版并开投机解码，速度约65 token/s。设备成本换算成API调用（GLM-5.2，每百万token 28元）可买约1178M token，需连续运行209天才能回本。512G版（108749元）运行量化GLM-5.2速度仅17 token/s，回本约7年。模型每1.5个月更新，建议普通用户买coding plan或租卡。已有Mac或显卡者，闲置时跑模型才划算。

Berryxia.AI@berryxia · 6月22日61

苹果换帅后会有什么大动作？ 这哥们本身就是硬件产品出生… 2001年，苹果发布了一款叫iPod的音乐播放器。没人想到它会重新定义整个消费电子行业。 二十五年后，苹果正在做一件更疯狂的事：把摄像头塞进AirPods。 Bloomberg的Mark Gurman报道，苹果计划在2027年推出三款产品：带摄像头的AirPods、第二代折叠屏iPhone，以及一款纪念iPhone诞生二十周年的特别版Pro机型。 其中最激进的是AirPods。内部代号B798，在耳机柄里放了计算机视觉摄像头。 看起来还是你熟悉的AirPods Pro，但它能看到你看到的东西。 苹果把这叫做"AI在身体上的下一个战场"。 这不是一个产品发布。 这是一整条产品线的重新定义。 耳机不再只是听音乐的工具，它变成了你和AI之间的视觉接口。 苹果在2001年用iPod证明了它可以重新发明一个品类。 二十六年后，它准备再做一次。只是这一次，品类是耳机。

译据Bloomberg报道，苹果计划2027年推出三款产品：内部代号B798的带摄像头AirPods、第二代折叠屏iPhone及纪念iPhone诞生二十周年的特别版Pro机型。最激进的是AirPods，在耳机柄内置计算机视觉摄像头，可捕捉用户视野，苹果将其定位为“AI在身体上的下一个战场”，使耳机从音乐工具变为AI视觉接口。新CEO John Ternus的首要任务被指是重整苹果设计团队，为产品线变革铺路。

Emad@EMostaque · 6月21日32

There will be an open source fable-level model that runs on a base MacBook mini / Air or equivalent. I don’t think people have realised this.

译将有一个开源寓言级模型，可在基础MacBook mini / Air或同等设备上运行。 我认为人们还没有意识到这一点。

Chubby♨️@kimmonismus · 6月19日56

I think many people still underestimate the full implications of China potentially gaining access to ASML-level EUV technology. The supply chain currently works like this: -ZEISS in Germany produces the ultra-precise optics and mirrors that are essential for ASML’s lithography systems. -ASML is the only company in the world that can supply EUV lithography machines at industrial scale. These machines use extreme ultraviolet light to expose tiny chip structures onto silicon wafers, enabling the production of the most advanced semiconductors. -The machines are shipped in many separate modules, then assembled and calibrated at fabs such as TSMC in Taiwan. This is where companies like NVIDIA, AMD and Apple have many of their most advanced chips manufactured. The key point is this: EUV is not just one machine. It is an entire industrial ecosystem. Optics, light sources, metrology, software, precision engineering, service teams and years of manufacturing know-how all have to work together. That is why there is no real alternative to ASML for leading-edge chip production at scale. This is also why China has been denied access to EUV for years. China has made enormous progress in AI chips and semiconductor manufacturing, especially through Huawei and SMIC. But it still lacks the mature industrial ecosystem needed to replicate ASML’s EUV capabilities at scale. Huawei’s Ascend chips are improving fast and are now estimated to reach roughly 60 to 80 percent of H100-class performance in some workloads, but they still lag behind NVIDIA in software, memory, reliability and large-scale training infrastructure. That is why the latest reports are so important. Reuters previously reported that a Chinese team, including former ASML engineers, had built an EUV prototype in Shenzhen. Now, the US government is reportedly concerned that China may have gained access to ASML’s most sensitive lithography technology. ASML denies that it has ever shipped EUV machines or EUV-specific components to China. If China really has gained meaningful access to EUV know-how, this would put the West under serious pressure. It would not mean China can suddenly mass-produce cutting-edge chips tomorrow. But it could shorten the timeline, accelerate Huawei’s chip roadmap and weaken one of the West’s most important technological chokepoints. Many people are missing the significance of this. EUV is not just about chipmaking. It is one of the deepest bottlenecks in the global AI race. (Image: ASML EUV)

译EUV光刻机依赖德国ZEISS光学、ASML整机及台积电组装构成的工业生态系统，中国被禁多年。华为Ascend芯片在部分负载中达H100级60%-80%性能，但软件、内存和训练基建仍落后。近期路透社报道前ASML工程师助中国在深圳逆向建造EUV原型，美政府担忧中国已接触ASML敏感技术，ASML否认。若中国真正掌握EUV，虽不能立即量产，但可能缩短芯片追赶周期、加速华为路线图，削弱西方在AI竞赛中的关键瓶颈。

OpenBMB@OpenBMB · 6月18日59

Really impressive “gauge reader” demo by @aijoey MiniCPM-V 4.6 👀 What makes this interesting is that it goes far beyond OCR: The model needs to understand multiple visual signals at once, including pointer angles, scale ranges, units, value mapping, digital displays, and liquid level proportions, often within the same scene.💥 This demonstrates strong visual reasoning ability, not just text reading 🧠 Even more importantly, the real-world setup matters here. Many factories, data centers, labs, and energy systems still rely on traditional gauges and legacy panels.👍In the industrial automation field, this will have huge application scenarios. Relying on MiniCPM‑V 4.6’s structured output and powerful multimodal capabilities, many traditional instruments without sensors can be retrofitted at low cost using this solution. 🔥Instead of replacing hardware or installing new sensors, this demo shows a practical path where cameras and vision models turn existing equipment into readable, recordable, and alarm-ready data sources. Big thanks to Joey for this great demo 🤝

译面壁智能 MiniCPM-V 4.6 演示工业仪表读取，模型需同时理解指针角度、刻度范围、单位、数字显示、液位比例等视觉信号，输出结构化 JSON（pressure_bar, temp_c, flow_lpm, level_pct）。测试使用合成控制面板，评分标准为 pass（满量程5%内）、drift（10%内）、miss。数字显示和液位较易，模拟指针更困难。该方案通过摄像头+视觉模型低成本改造传统仪表，无需更换硬件，在工厂、数据中心等场景有巨大应用潜力。

Berryxia.AI@berryxia · 6月17日65

兄弟们，肝出来了！并且全部免费开源了！ 我把PP-OCRv6直接做成了本地工作台，在Mac上用CoreML加速，一键就能切换Tiny、Small、Medium三个模型大小！ Tiny只有1.5MB适合极致轻量，Medium 34.5MB主打精度，中间Small做平衡。 支持图片上传、批量处理、结果导出CSV/Markdown/Excel，历史记录自动保存。 整个东西完全本地跑，隐私安全，不用上传任何数据。 最爽的是在苹果硅上自动开启CoreML加速，Intel Mac和Linux也能CPU跑。 还专门做了浏览器版Tiny模型，零依赖直接在网页里就能用OCR。 附带评测脚本，能跟OmniDocBench和macOS自带Apple Vision对比，实际测试在弯曲表面、点阵字体、低对比度这些难搞场景表现都不错。 以前做本地OCR最烦的就是模型下载、环境配置、精度和速度权衡，现在我把这些全封装好了。 开发者、研究者、需要离线处理文档的人直接clone就能用。 这其实是我自己日常做OCR时踩了很多坑后，顺手做的东西。 希望能帮到同样有这个需求的朋友。

译Berry Xia 开源了基于 PP-OCRv6 的本地工作台，在 Mac 上使用 CoreML 加速，提供 Tiny（1.5MB）、Small、Medium（34.5MB）三个模型大小，支持一键切换。支持图片上传、批量处理、结果导出 CSV/Markdown/Excel，历史记录自动保存。全部本地运行，隐私安全；苹果硅自动开启 CoreML 加速，Intel Mac 和 Linux 也能用 CPU 跑。另提供浏览器版 Tiny 模型，零依赖即可在网页端使用。附带评测脚本可对比 OmniDocBench 和 macOS 自带 Apple Vision，在弯曲表面、点阵字体、低对比度等场景表现良好。