Hy 3D WorldClaw is out🤯 An agentic workflow that turns a text prompt into a large scale 3D open world. Every scene is freely explorable🌻, built from editable, game-ready 3D assets with real geometry and textures. actual meshes, not video or splats. Project page 👉

We went in assuming the harness was just a ruler, turns out it isn’t. We found this the hard way, trying to answer an internal question: which model for which kind of work No eval covered it, so we built one. WorkBuddy Bench is open now, 260 tasks across code, web, office and security. Three things we didn't expect: — no model won everything. The leader changed track by track, and across the models we ran, GLM-5.2 came out on top for security on both setups we tried — the scaffolding around the model moves scores as much as the model choice does. Same model, same tasks, different client, and the security ranking reshuffled — real coding tasks are hard not because of the code, but because of the context. In our Code subset, bug fixes and API contracts were the toughest categories We don't think this settles anything, real work is messier than any 260 tasks can capture, and that's exactly the part we want to keep working on. If you've got a read from your own practice, we're listening repo👉🏻

🗄️PostgreSQL just landed in CloudBase With CloudBase, your Vibe Coding product can spin up its own database with a single prompt In practice: - Tables, permissions, and auth rules are declared in SQL, the same language your Agent already reads - Row-level security means author_id = auth.uid() does the heavy lifting, no custom middleware - Each user environment is isolated by default: spun up, hibernated, or deleted via API (or MCP) Go build→：

It finally happened: instead of a single 3D asset, AI now generates an entire 3D world full of editable assets. Tencent just dropped WorldClaw - an agentic system that builds and refines full environments inside Blender from one prompt. • Region-aware procedural terrain • Independently editable textured meshes • Automatic object generation + placement • Agents fix scale, pose and terrain contact • Scene construction and refinement in Blender Still very early and far from perfect, but this is a significant step toward actual AI-generated 3D worlds. Source:

FlexKV came up the most in yesterday's replies, so here's the longer version： It's a KV cache layer that sits under your inference engine, open sourced last year. works across SGLang, vLLM, TensorRT-LLM and Dynamo. If you serve long contexts you know the shape of this: cache fills VRAM, gets evicted, and you re-prefill a prefix you already computed once. A cache is only worth what it hits, most of FlexKV's design goes into keeping that number high: — maximizes effective cache capacity through cross-rank deduplication and lossless nvCOMP ANS compression, across CPU, SSD, Mooncake Store, and GDS — maintains a cluster-wide distributed RadixTree for cross-node prefix discovery and reuse — exposes cache residency and prefix matches to Dynamo for cache-aware routing DeepSeek-V4 support landed last month too, heterogeneous KV groups and compress-state sidecars included, with layerwise loading to cut read latency. SGLang path is ready now, other engines coming Repo→

Y'all know Valkey right? We co-maintain it upstream And now we support Valkey 9.0 (up to 200% throughput gains on compute-heavy ops) with a set of kernel-level patches built for prod: - Stateful connections (Pub/Sub, MULTI) survive node migration without dropping - Rehash latency spikes minimized at the root, keeping data service stable and smooth under load - Same-zone replica wins the failover election, avoiding the latency hit and performance loss from cross-zone switches For agent builders: pipeline memory prefetch in 9.0 cuts batch read latency by up to 40% and zero-copy responses add another 20% on large value reads. I bet it's worth a look:

Open weights are table stakes now So here's the boring infra nobody wants to tweet about, also open⚡️ Things we've open sourced: — CubeSandbox, per-conversation sandboxes with browser, code, shell and file access, spun up and torn down for you — Agent Memory, a shared memory layer across your agents, and now across your teammates' agents too Things we've upstreamed: — RoCE optimizations for DeepSeek's DeepEP, doubling throughput on commodity networks, with 30% carrying back to InfiniBand — FlexKV, tiered KV cache offloading, now in NVIDIA Dynamo, vLLM and TensorRT-LLM mainlines All four are searchable by name, repos and docs. Pull requests welcome, and so are complaints 🛠️

🤯Introducing Team Memory, same idea as Agent Memory, except your teammates' agents can read it too 2.0.0 beta out today, and the repo hit #1# on github's typescript trending this week Highlights: > Solo builders: one place to manage memory across all your agents and AI tools, chat, code, tasks. Built for the one-person company > Teams: a shared memory hub that turns conversations, docs and code into four reusable assets, Chat Memory, Skill, LLM-Wiki, Code-Graph, governed and shared across agents and frameworks changelog and repo →

A month after we shipped Hy3, the more interesting question became where you can actually use it More entry points live now: >WorkBuddy, our agent workspace. In internal evals at tencent, Hy3 delivered a task success rate above 90% and cut average task time by 34% over the previous Hy model. Grab it at >Miora, our creative studio, Hy3 handling reasoning across multi-agent design work >TokenHub, Tencent Cloud's MaaS platform, with routing across Hy3 and other models More of the Tencent Cloud AI-native stack is next, and the API and open weights aren't going anywhere what's the touchpoint you'd want to see Hy3 in next?

🌍Introducing Hy ASR 3.0 preview, a speech recognition model from @TencentHunyuan built to understand, not just transcribe What we improved: - Cleaner on real-world audio: dialects, code-switching, and long-form content with less error accumulation over time - Context-aware correction: homophones and ambiguous phrases get resolved from context, not guessed in isolation - Hotword injection: drop in brand names, people, domain terms without retraining. Lowers integration cost for niche use cases - Built for noisy rooms: whisper, background noise, tricky acoustic conditions stay stable

🧬 Over the past few years Tencent has gone deep on AI for Life Science. We've built single-cell models, AI-designed antibodies, and now cell-level drug prediction. Today we share one of our latest results: UniPert-G2CP, a two-stage framework (published in Cell) that predicts how a drug molecule affects a specific cell, trained and evaluated on 4,994 genes × 7,860 compounds across 5 cancer cell lines. I'm not a specialist, so here comes the easy version: the same drug can work very differently across genetic backgrounds and cell types. Prior research mostly worked at the molecular level, but our result gives AI the ability to predict how a drug molecule affects a cell. For the real mechanics, the full Cell paper is here👀: We'd love your takes

Something we have been exploring: AI that co-writes the document with you, not just chats about it 🔥 We built it into WorkBuddy, our AI agent for work. It edits the same file you are working in, so it just sees your edits instead of being told, and it never overwrites them. Local Word, Excel, PPT, Markdown, and Tencent Docs online🙌

Zen garden for $0.65 🤩 The details carry it: light the lanterns, scatter koi food, ring the great bell, rake the sand. Everything in it is interactive, at about 1% of the top-of-chart cost. Thanks to everyone pushing Hy3 this far

Open source update 🤯 CubeSandbox v0.6.0 is live with Kubernetes and Volume support, led by these two community requests: 🔹 Cube no longer needs a separate sandbox cluster. It can be deployed, scheduled and operated like a regular K8s workload. 🔹 Bring your own storage to Cube. The E2B-compatible Volume framework uses Create, Destroy, Attach and Detach hooks to decouple sandbox storage from any specific backend. 👉:

AI-Infra-Guard is an open-source, self-hosted red-teaming platform for AI builders. • Scan AI infrastructure for fingerprints and CVEs • Audit MCP servers and Agent Skills • Red-team Agent workflows and LLMs • Run locally with Docker; integrate via Web UI or API Build with more visibility:

Miora is now fully live! Our AI-native studio for creative work. Huge thanks to everyone who beta-tested with us. Two new things shipping with the launch: — Multi-agent collaboration: a main agent orchestrates specialized sub-agents for different parts of a design task — Agent memory: remembers not just the project, but your style and taste New here? Get free 1,000 points when you sign up, then come create with Miora! 🔗：

✅Refreshed SOTA on 29/55 open math problems ✅10-digit addition transformer with just 15 trainable params (58.3% fewer than public record) ✅Sunspot prediction with r² = 0.77 on nearly a century of held-out data ✅44.4% more efficient qubit routing than SABRE on IBM Q20 ✅Candidate parp1 inhibitors that outscored olaparib in silico（early-stage, wet-lab work still ahead） 🔗All outputs open-sourced →

Robot Xiaoliu clocked in at #WAIC# 2026, and the job today? Giving massages.🤖😜 This year, embodied AI and real-world agents have taken over the show. Today we're launching our full-stack embodied AI solution, moving from disembodied to embodied, from fragmented modules to one closed loop, across four layers: → Models. The Hy-Embodied matrix: RxBrain-1.0 pairs text reasoning with visual imagination, VLM-1.0 hits flagship perception at 1/10 the compute, VLA-0.5 unifies vision, language, and action. → Agent framework. TairosAgent and Apexio integrate the brain, cerebellum, and body into one system, on the upgraded Tairos platform. → Cloud infrastructure. From GPU/HCC compute and Astral Network to TI-ONE, TokenHub, and TRTC/TRRO, covering training, serving, and real-time interaction. → Industry applications. The industry's first cloud-based EaaS, Embodied-AI-as-a-Service, taking embodied AI from breakthrough to scale.