We’ve integrated @TencentHunyuan's AI-Infra-Guard (AIG) into ClawScan, the open-source command-line tool that powers security review on ClawHub. Every skill and plugin uploaded to ClawHub now runs through AIG as part of its security review. openclaw.ai/blog/tencent-aig…

Thursday. Our editor’s still on vacation, so today’s temp crew is nine Tencent penguins — all generated by the new Hy Image3.5 preview. they’re hiding,help us catch them. 🕵️🐧

Any image model can nail one pretty shot. The hard part is holding the brief through every revision after that. Serrino from @Miora_Design says that's what professional grade means for Hy Image3.5 preview. His take is that the model raises the ceiling and the product is how you actually reach it. We think that loop is where a model gets good, with real work feeding back into what it needs to fix. Full episode with @Gracemzshao on AI Proem, a Behind the Penguin collab: open.spotify.com/show/6qsAjS…

The most common break in AI video: the same character, a slightly different face in every shot. 🎬 At @AsianFilmAwards, our @Miora_Design AI product expert walked through our fix: → every character gets 3 reference sheets — the front close-up is the identity anchor → every generation reads the same sheets, not the model's memory → the sheets do the remembering, so the character doesn't drift Consistency isn't a model problem. It's a pipeline problem. What still breaks consistency in your AI videos? 👀

open-source update: you can give a new OpenClaw Skill a quick safety check with one line in chat. ClawScan is AI-Infra-Guard's module for this. It looks for potentially malicious Skills, exposed credentials, and known CVEs, then returns a report. • AI-Infra-Guard was selected for Black Hat Arsenal • its vulnerability database covers 2,000+ CVE rules across 146 AI components, including Ollama and ComfyUI github.com/Tencent/AI-Infra-…

🚀 Tencent Hy Translation just landed.Powered by Hy-MT2. 33 languages. 5 Chinese minority languages & dialects. Voice. Photo. Full offline — on-device, no network required. Already live in 12 countries and regions. Travel, drive, work or read abroad. Accurate. Natural. Always available. ⏬ ⏬⏬ apps.apple.com/us/app/tencen…

Your agent's chat window is the whole pipeline now. Quick one: WorkBuddy in China now takes mini programs from description to published. Publishing is the part that changed today: > one QR scan authorizes WorkBuddy as a third-party publisher on your account > upload, trial build, review and release all run from there Mini programs being the lightweight apps inside Weixin, nothing to install. The mini program arrives with database, Weixin login, phone login and file storage already in it. The ecosystem part is what we keep coming back to, fewer steps, more builders.

Two weeks since the TeamAI post, and the same handbook covers a lot more agents: > kiro, github copilot and oh my pi joined this week > rules, agents, mcp servers and hooks, scoped by role and project > learnings on their own branch now yours not on the list? that's what PRs are for. github.com/Tencent/teamai-cl…

Interesting take on world model consistency: what if you skip 3D reconstruction entirely? WorldCrafter from TencentARC, weights are open. Instead of building 3D geometry, the video generator queries an implicit memory by camera pose. point the camera somewhere and the memory fills in the view. → one image or text prompt, minutes of scene exploration. base and distilled fast version both on HF → 384×640 for now. resolution can scale once the architecture works ✍️full paper at arxiv.org/abs/2609.24984 🤗weights at huggingface.co/TencentARC/Wo… ⌨️code at github.com/TencentARC/WorldC…

A professional image model that doesn't cost like one. $0.024 per 2K image. Meet Hy Image3.5 preview. > up to 5 reference images in, and the meter only runs on what comes out >multi-turn editing with earlier edits in context > text, layout and 2K consistency, the parts that break posters and UI mocks Free for two weeks in @OnSoloAI and @Miora_Design . API:console.tencentcloud.com/tok… Previews exist to get corrected, tell us what's off and it goes into the next version.

Close your laptop. Keep your AI agent working. This 3-step guide shows how to run OpenClaw on Tencent Cloud Lighthouse, send tasks via WhatsApp, and set up overnight price and inventory checks—with a morning report on your phone. 👇

Search it. See it. Keep reasoning with it. 👀 That last part is harder than it sounds. In many search-agent setups, images retrieved along the way are not preserved in the model’s later context. As a result, a seemingly multimodal search trajectory can gradually collapse back into text-only reasoning. WeAgent-MMSearch, developed by the Weixin AI team at Tencent, is designed to keep visual evidence in the loop. With a persistent multimodal state, retrieved images remain accessible throughout the search trajectory. The agent can revisit them, cross-check visual evidence across turns, perform reverse-image searches, and reason over the retrieved images directly in subsequent reasoning and tool calls — rather than reducing everything to text. With agentic SFT followed by Failure-Aware GSPO, WeAgent-MMSearch reaches an average score of 55.97 across 8 benchmarks, up 19.22 points from baseline. We also introduce VisTarget-Bench, a human-verified benchmark of 150 tasks designed to disentangle two failure modes that standard evaluation often conflates: failing to retrieve the right visual evidence, and failing to answer correctly even after that evidence has been retrieved. Each task is paired with a held-out target image, allowing us to separately measure whether the agent retrieves the intended evidence and whether it can answer correctly once that evidence has been found. For multimodal search, finding the evidence is only the beginning. More on WeAgent-MMSearch: kkkaiaiai.github.io/WeAgent-…

This might be the first open-source memory architecture to cover associative recall. The rest stop at similarity search. (T-Mem, EMNLP 2026) Most memory systems retrieve by similarity. T-Mem predicts when a memory will matter, before you ask. → write: memories ship with a trigger at store time, basically "when will this come up again?" → read: at query time, triggers pull in memories by situation, even with zero keyword overlap. → architecture layer. plugs into any LLM without fine-tuning or model changes. On LoCoMo-Plus (strips keyword overlap, tests pure associative recall), mainstream systems drop 28 to 50 points. T-Mem drops 5.45. paper: arxiv.org/abs/2606.15405 code: github.com/Sherlockwz/T-Mem (MIT)

The demo runs on your GPU now. It wanted ~25GB of VRAM on launch day. 💻This week: CPU offload, a ~7.5GB encoder cut, an MLX branch for apple silicon.

Tencent just dropped a model that edits audio content, voice, and emotional tone from a single instruction. Same words. Different everything else.

We open sourced BrowserSkill, a bridge between your agent and your actual browser. most tools give the agent a blank browser. We let it borrow a tab from yours, then hand it back. > login state is already there, it just works where you're signed in > captchas and confirmation dialogs come back to you, then it continues > it's a CLI, not an MCP server => any agent that can run a shell can use it, and you see every call it makes one thing that's easy to miss: the agent asks before borrowing a tab, and that switch lives in your browser settings, not in a flag, so it can't be talked around. one line to install, works with Cursor, Claude Code, Codex, Hermes, Openclaw, CodeBuddy, WorkBuddy. Everything runs locally, MIT. github.com/Tencent/BrowserSk…

朱啸虎对腾讯 AI 的真实评价。 朱啸虎昨天的新访谈，有几个观点我还是比较认可。总结下： 第一，模型 API 这门生意，很难长期维持高毛利。 Anthropic 现在能做到 60% 左右的毛利，核心还是模型智能能力比较强，尤其 Coding 这类场景，过去一段时间确实有明显优势。 但这个优势已经开始越过临界点了。当其他模型逐渐追上，价格可能只有你的十分之一，怎么可能一直维持这么高的溢价？ 哪怕 Anthropic 的模型未来几年还能保持领先，但如果智能能力只强 10%，价格却高 50%，用户真的会一直买单吗？ 智能最终会越来越像电和水。模型 API 这一层，长期可能就是 10% 到 20% 的毛利。这是比较合理的。 一旦能力逐渐趋同，价格战很难避免，大家也会越来越有动力切换到更便宜的模型。Anthropic 现在着急上市，也可能和这个窗口期有关。 第二，AI 办公的市场肯定比 AI Coding 大很多。 AI Coding 最先跑出来，很重要的一个原因是工程师本身就对 AI 的接受程度就比较高。而且 Coding 是一个特别适合 AI 的场景，任务相对明确，结果也比较容易验证。 但 AI Coding 只是办公的一部分。真正更大的市场，还是整个的白领市场，也就是现在大家都在拼的 AI 办公。 再进一步，AI 办公本质就是通用 Agent。之前的 Coding 只是一个垂直场景，办公是整个白领市场。就像当年微软的 Office 一样，你会发现只要有电脑的地方，就会有 Office。 第三，AI 办公产品的竞争力是 Model + Harness + Product + Context。 今年前半年，行业聊 Harness 比较多。但 100%可以预测到，随着 Harness 这一层趋同，最终的竞争肯定会转移到 Context。 因为 AI 办公的杀手级应用场景是工作，诸如分析数据，写 PPT 和分析报告等等，这类任务的交付结果，直接取决于 AI 到底掌握了多少企业的工作上下文。 就像一个新员工一样，你想让它把工作干好，又不让它了解背景，那即便他再聪明也无济于事。 Context 肯定会成为这一轮 AI 办公竞争中非常重要的壁垒。 基于这三个判断，朱啸虎也聊到了腾讯的 AI 发展。相对还是比较客观。 当然，我也不否认朱啸虎有他的局限性，大家主要看看他的思考逻辑，是不是对我们理解一家公司和趋势有启发。 腾讯在这一轮 AI 浪潮里，目前看还是非常典型的后发制人。 最开始混元模型，说实话，在业界都没什么存在感。但现在做到 Hy4 preview，已经逐步的挤进国内第一梯队了。 到了这一轮 AI 办公，WorkBuddy 反而跑出来了，而且已经成为国内第一梯队的办公 Agent 应用。 朱啸虎说金沙江也购买了 WorkBuddy 的企业账号。 为什么？ 一方面当然是 WorkBuddy 本身的产品体验不错。另一个很重要的原因是金沙江本来就在用企业微信。 企业微信里已经沉淀了公司的组织关系、历史沟通，再加上腾讯会议、腾讯文档，本身就在同一个体系里。对于 WorkBuddy 来说，这些都是现成的 Context。 如果换成一个独立的 AI 办公产品，就要重新连接知识库、文档、会议，甚至重新处理组织关系和权限。 朱啸虎还有一个判断我也很认可。现在模型能力当然还重要，因为大家还处在交替领先的阶段。但只要模型之间的差距继续缩小，技术优势就很难成为长期壁垒。 到了那个时候，竞争又会回到传统互联网的几个指标上。比如谁掌握了入口、谁有获客优势、谁能拿到用户的历史数据。这些恰恰是腾讯这样的公司最有优势的地方。 如果把去年的 ChatBot、AI Coding，再到今年的 AI 办公放在一起对比，会发现 AI 应用的竞争焦点确实一直在往上层迁移。 最开始的 ChatBot 拼的是模型能力，那时候谁的模型能力强，谁的产品体验就更好。但到了后来的 AI Coding，大家发现，光模型强还不够，Harness 也很重要。因为在长程任务里，模型怎么调用工具，怎么处理错误，会直接影响最终的结果。 现在到了 AI 办公，又开始发现，Model 和 Harness 上面，还有 Context。 所以我一直觉得，AI 办公这种通用级 Agent，最后大概率还是大厂或者超级明星创业公司的生意。因为做到这一步，已经很难靠一个单点能力赢了，后面牵扯的东西会越来越多。

Can we get a redo button while we're at it? Secure VM put agent isolation on everyone's mind. Rollback is the other half of safety. When a step goes wrong, you want to undo it. @CubeSandbox is open source. Agents go off-script all the time, so each task gets its own kernel and untrusted code never touches your stuff. Your credentials sit in a vault and get injected at the gateway, never inside the sandbox. And when a step still goes wrong, you roll back to a snapshot instead of starting over. 🤯repo: github.com/TencentCloud/Cube…