Feng Zhou (@fengzhou) profile picture
Beijing
Joined May 2008
Photo by Feng Zhou (@fengzhou)Photo by Feng Zhou (@fengzhou)Photo by Feng Zhou (@fengzhou)Photo by Feng Zhou (@fengzhou)Photo by Feng Zhou (@fengzhou)Photo by Feng Zhou (@fengzhou)Photo by Feng Zhou (@fengzhou)
5 days after release, both R2T2 (2B, live transcription) and T3PO (14B, text-to-text simultaneous translation) rank #1 in their Hugging Face categories🚀 High-quality open-source multilingual transcription and translation from Youdao. Can’t wait to see what you build with them.
Meet Confucius4-R2T2: a Qwen3-ASR 1.7B based, low-latency streaming transcription model ⚡ Streaming chunks as small as 80 ms 🎯 SOTA WER/CER for Chinese & English among open models 📝 Stable, append-only transcripts Github: github.com/netease-youdao/Co… Try: huggingface.co/spaces/huggin…
1
2
4
482
Meet Confucius4-R2T2: a Qwen3-ASR 1.7B based, low-latency streaming transcription model ⚡ Streaming chunks as small as 80 ms 🎯 SOTA WER/CER for Chinese & English among open models 📝 Stable, append-only transcripts Github: github.com/netease-youdao/Co… Try: huggingface.co/spaces/huggin…
2
3
2
7
1,389
Model architecture of Confucius4-R2T2, featuring an LLM decoder trained with our original Longest Stable Prefix (LSP) objective.
1
143
Confucius4-R2T2 achieves strong streaming ASR accuracy at very low latency. At 160 ms chunks: ⚡ 6.13 WER on English, 6.42 CER on Chinese 🎯 Best accuracy among the true-streaming open models we tested 🚀 R2T2 also offers a latency-priority mode for even faster responses.
68
🚀 Just released: Youdao Confucius4-TTS An open-weight 1.3B high-quality TTS engine: - 14 langs: EN, ZH, JA, KO, DE, FR, ES, ID, IT, TH, PT, RU, MS, VI - High-quality voice cloning w/o transcript - Cross-lingual voice transfer with min accent - Emotion transfer - Apache License
4
8
3
17
33,592
Feng Zhou retweeted
LobsterAI 2026.6.1 Release 🦞 ✨ New Features & Models • Expert Kits: 1-click install grouped capabilities tailored for your workflows! • Fork Conversations: Easily branch out your chat history to explore new ideas. • Support MiniMax M3 and optimized context windows. ⚡ UX & Performance • Lightweight collapsible previews for massive outputs—no more page lag! • Clearer token consumption info for Video Gen & new account quick links. 🛠️ Under the Hood • Smoother MCP runtime cleanup & fortified AI engine stability. • Fixed WeChat login state loss on Windows updates. github.com/netease-youdao/Lo…
1
1
1
256
Thanks for the mention🫡Appreciate the love! If you haven’t tried LobsterAI yet, give it a spin — we’re evolving fast.
14
7
18
61
292,987
WeChat integration is Huge.
#LobsterAI v2026.3.22 is LIVE with full #WeChat integration.🦞💬 Tired of being chained to your PC? Stop juggling apps and start using your favorite AI Agent directly inside WeChat. 💥Access anywhere: Ultimate flexibility for remote work and on-the-go tasks. 🧠 Zero learning curve: Just send commands in your chat box—super intuitive. ⚡️ Instant execution: Get task results without leaving your conversations. Put an ultra-efficient AI in your pocket now~ #AIAutomation #AI
2
686
Feng Zhou retweeted
Exciting Update for #LobsterAI! 1⃣ Skill Marketplace:Want new features? Just one click~😀 Think of it as Lobster's app store. Explore ready-made "skills" like data analysis, browser helpers, note syncing, and more. No hassle—install directly from the Skill Marketplace! You can even add community skills or upload your own. Ready to use with just a command! 2⃣ MCP: Connect Lobster to your favorite apps~🤩 Your personal Lobster can now work directly within tools you already use—Notion, GitHub, browsers, and collaboration software. It can read and process data right from your apps. 💡 LobsterAI is now fully compatible with both macOS and Windows! Download from our website and join developers on GitHub to help us build.❤️ Whether you’re a power user or a developer, #LobsterAI is designed to make your AI work for you, not just with you. #AI #LobsterAI #Productivity #OpenSource #MCP #TechUpdate #LLM
1
3
1
8
2,176
Feng Zhou retweeted
LobsterAI V0.1.24 Release 🦞 📐 Support formula rendering 🐝 IM supports NetEase Bee 🖼️ Supports passing images to vision model 🐧 Linux client packaging optimization 🔧 Fixed bugs such as email connection issues on the Windows client github.com/netease-youdao/Lo…
5
1
6
1,217
Ollama experience is improved. Enjoy.
2
5
1,215
LobsterAI v0.1.21 Release 🦞 ✨ New: • Kimi/GLM/Qwen/Volcengine coding • Media input (TG/Discord/DingTalk/Feishu) • Custom proxy toggle • Volcengine provider 🛠 Fixed: • Skills parsing & MiniMax format • Windows path (Chinese/git-bash) issues github.com/netease-youdao/Lo…
2
2
4
997
Feng Zhou retweeted
昨天,网易有道悄悄开源了一个非常有意思的国产 AI Agent —— LobsterAI 。 体验完后最大的感慨:这简直就是带 GUI 的中国版 OpenClaw,但门槛被拉低到了普通人也能轻松把玩的程度。 现在的 AI 产品有两个痛点:要么纯云端(常遇延迟和断网),要么纯聊天(没法进行深度的交互和操作)。而 LobsterAI 是一个基于 Electron + React 的本地优先 Agent。你只需用自然语言聊天,它就能直接在本地电脑上操作你的文件系统和终端, 同时支持最全的主流国内外的聊天软件(飞书,钉钉,Telegram,Discord)去远程控制,比别的工具都丰富,内置 memory 长记忆,会记住你的偏好,跨 session 自动复用。 这是官网,可以试试: lobsterai.youdao.com/#/index 结合我自己的需求,直接用它跑通了 3 个刚需场景: 1 每天手动刷 HackerNews 并且找有趣的内容比较浪费时间,我让它给我每天早上九点精选 HackerNews 上有趣的十篇文章进行摘要和翻译,生成 Word 文档,然后每天定时发送到我的指定邮箱,这样吃早餐的时候就可以浏览汇总的文章了。 2 简单简历一直在更新最新的简历范本,也是网站 SEO 很重要的一部分,之前需要脚本的方式,经常会出现问题,现在我让它按照每天一次按照热度生成不同的简历职位,然后生成这些职位的简历范本数据,用文件持久化进度,同时将简历数据直接插入到本地数据库中,在我确认之后直接推送到线上数据库。 3 我用手机钉钉上让它调研了一个出行计划并且保存在我本地的 Obsidian。它在家里电脑上自动:浏览器搜索攻略/比价 → 整理成表格 → 生成Obsidian笔记 → 同步给我手机。手机随时指挥电脑,真正的跨端私人助理。 远程控制做的比较好:支持飞书、钉钉、Telegram、Discord 远程操控,手机随时指挥电脑。 内置长记忆 (Memory):能记住你的已有信息和偏好,跨 session 自动复用,越用越顺手。 定时任务 & 沙盒环境:支持每天定点执行(极其实用);对数据敏感可以开启沙盒,防止 AI 暴走破坏本地数据。 多模型兼容:支持市面上所有主流 Provider,跑本地模型(比如 Ollama)也没问题;内置丰富的 Skills,做视频、幻灯片、写文档全包揽。 大年初三刚开源,整体体验下来完成度挺高,集齐了各种功能,本地化做得极其贴心。受够了纯聊天 AI、想让 AI 真正帮你操作电脑“干活”的,强烈建议体验一下。 #AI #LobsterAI #Agent #youdao #netease
35
221
5
897
169,936
We’ve found VM is also good for compatibility under Windows
Replying to @fengzhou
persistent memory + skills system is the right architecture. been running a similar setup and the VM sandbox is smart for trust. curious how the IM integration handles multi-step workflows
606
界面语言设成英文试下
Replying to @fengzhou
你好,为什么我这个模型列表里没有gemini
1
1
1,159
大小性能这些Qemu问题都不大,也带加速,看一段时间,目前Claw类技术方案还没定型。
Replying to @fengzhou
为什么使用 qemu 虚拟机作为沙箱环境?是不是有点重?
1
625
的确一个目标就是对新手友好,MCP这些在计划中。
试用了下有道的 Claw 分身 LobsterAI,简单说,很适合新手用,可以添加 Skill,但要注意,没有 MCP,无法外接记忆系统,无法设置多代理和子代理,没有 Agent 或者 Soul 文件,也就无法设置角色。 总结:要不试一试 Alma,比这个好玩。
7
1
3
2,819