Tencent AI
@TencentAI_News
The official @Tencentglobal newsroom for AI updates and developer resources.
96 Following 25.5K Followers
Hy-Nomas now runs on two more backends.
SGLang (main branch), by @sgl_project : offline video jobs and streaming sessions.
ComfyUI, by @LatentSpacer : chained segments and WASD camera control.
Thank you both for building on it.
Show more
🚀 Introducing Hy-Nomas: An Open Source Free-Form Interactive Video World Model
Key Features:
* Free-form interaction with text prompts and camera trajectories
* Long-horizon generation with world memory
* Lightweight LoRA adaptation for specialized domains such as robotics
* Optimized for Inference on Consumer GPUs
Test it and drop feedback anytime！
🔗 Website:
💻 GitHub:
🤗 Hugging Face:
🎥 Demo:
#Hy##WorldModels# #VideoGeneration# #GenerativeAI#
Show more
You can now curl a simultaneous interpreter.
Hy-SimulMT runs on your own GPUs and streams Chinese and English speech translation as people talk. On the open RealSI long-form benchmark it holds lag to about 3 seconds, and it scores 84.7 on RealSI-Sentence.
Six checkpoints are out under Apache 2.0, with FP8 and GGUF versions among them. The repo ships an inference service plus training tools. Add a TTS service, and it answers in the speaker's own voice, in Chinese or English.
Put it on your audio and tell us where it drifts.
Repo:
Weights:
Show more
🎮 We’re open-sourcing Multics: Generative Multi-Player World Model with Hierarchical Communication!
Generating videos for multiple players requires each player to follow their own controls while staying consistent with others. As player counts grow, efficient communication and robustness to interference become key challenges.
Existing approaches face trade-offs:
• Fully-connected communication incurs quadratic attention costs as more players join.
• Centralized communication mixes all players’ features in a shared representation, causing interference and competition for communication capacity.
Multics introduces hierarchical communication across player, group, and global levels. Group-level representations aggregate information from related players, while global tokens enable cross-group communication. Shared group-token parameters let the same communication module support more players—without retraining or adding parameters.
Two key highlights:
✨ Robust multi-player generation: improved cross-player consistency, player-specific action and camera control, and interference robustness.
📈 Train on fewer, generate for more: trained with 4 players, Multics generalizes to 6, 8, and 10 players while keeping attention costs manageable.
Paper, code, and model weights are now available!
🌐 Project:
🤗 Paper:
💻 Code:
📦 Model:
Show more
The local bar for Hy Memory just dropped.
Size: 60 GB to 18 GB.
Stack: vLLM to one llama.cpp binary.
Quality: Q6_K 58.8%, Q4_K_M 56.5% on our LoCoMo eval.
Your memory bank stays on your disk.
🤗
Show more
A 3B active model beats Opus 5 at memory. We open-sourced the full stack — model, engine, and app.
Hy Memory handles chat, work, and code memory in one layer, and it's first on all four benchmarks we ran, leading MemoryArena by 10.2 points while cutting memory volume 51.8% and recall tokens 39.3%.
The model isn't tied to our engine. Drop Hy-Memory-30B-A3B into Mem0, MemOS, or EverMemOS and it still comes out on top, even on the one system it never trained on. It's already running in our own products, where one assistant's multi-hop recall went from 50% to 90%. It also unifies memory across Codex, Cursor, WorkBuddy, and more.
Repo:
Weights:
Questions welcome in the replies.
Show more
Your team's AI can now live on more than one machine, and every person gets their own.
It's Octop, our open-source, self-hosted AI workspace, where each person has separate agents, memory and workspace. The new 1.0.2 is about the team side, from how agents work together to where the work runs.
Team mode lets you describe the outcome once, and several agents figure out who does what before returning one answer. Cross-machine collaboration lets your laptop use another machine's experts, files and environment. Nothing moves, including conversations.
We also support decision models like Jev and Laya. Before a task reaches the main model, they screen which tools, skills and teammates are actually relevant, which cuts token use as the team and the toolset grow.
Use octop run to start the 10k-star project:
Show more
Thanks for sharing! Excited to see what your tests turn up. More open source where that came from.
An open model fully dedicated to memory management by @TencentHunyuan Hy-Memory-30B-A3B.
Really interesting! Bookmarked for future tests.
A 3B active model beats Opus 5 at memory. We open-sourced the full stack — model, engine, and app.
Hy Memory handles chat, work, and code memory in one layer, and it's first on all four benchmarks we ran, leading MemoryArena by 10.2 points while cutting memory volume 51.8% and recall tokens 39.3%.
The model isn't tied to our engine. Drop Hy-Memory-30B-A3B into Mem0, MemOS, or EverMemOS and it still comes out on top, even on the one system it never trained on. It's already running in our own products, where one assistant's multi-hop recall went from 50% to 90%. It also unifies memory across Codex, Cursor, WorkBuddy, and more.
Repo:
Weights:
Questions welcome in the replies.
Show more
Every production pipeline charges a tax nobody invoices: re-prompting your own idea at every step.
Miora @Miora_Design just removed it: story, cast, shots, final cut, one canvas.
Announced by Miora itself ⬇️ we had no notes.
Show more
Our WorkBuddy meetup at #SFTechWeek# brought some incredible workflows to life! From automated hiring pipelines to cinematic soundtrack generation and multi-platform content engines, our speakers showed just how far you can take WorkBuddy.
Good ideas are meant to be shared. Watch the highlights!
Show more
Meet LightVela, a 24/7 cloud AI assistant that also lives in the apps you already use, Telegram, WhatsApp, Discord.
Create a bot in seconds. Nothing to configure. Or create a few: one for work, one for learning, one for the kids. Each keeps its own memory and its own schedule.
Your laptop can be off. LightVela keeps the lights on.
First month free, with 4,500 credits included. Come break it:
Show more
Everyone's posting about RSI. So we built the exam AI can't study for.
Most benchmark questions have decades of human discussion behind them, and all of that text sits in the training data. So a high score might just mean a good memory.
We created two alien worlds where nothing a model learned applies: print 100 and you get 127 back, and the manual is wrong on purpose. Ten frontier models went in. None scored above 16% before exploring, and four rounds of experiments later, the best hit 89%.
Every score on the board is a measure of discovery.
Run the demo worlds yourself:
Show more
New Research: We are releasing ExplorationBench, a benchmark for measuring how AI systems explore.
Scientific discovery begins where known problems end: a system has to frame hypotheses, design experiments, and learn from the results. Evaluating this is hard. Genuinely new answers cannot be checked quickly, and in familiar domains a model can simply recall what it has seen.
Addressing this challenge, researchers from Tencent Hy, Fudan University, and Tsinghua University built verifiable Alien Worlds. Their rules are executable, so every answer is checked exactly, and they conflict with familiar knowledge, so recall alone cannot solve the tasks.
🔹 Two sandboxes: AlienCode (31 hidden rule changes, 70 tasks) and AlienLogic (24 patched inference rules, 70 theorems)
🔹 A flawed manual, four rounds of self-designed probes, and closed-book tests after every round
🔹 Every answer graded by an interpreter or a proof checker, with no LLM judge
What we found across 10 frontier AI systems:
1️⃣ Getting feedback is more effective than thinking alone. No AlienCode run starts above 15.7%; after four rounds the best reaches 89.0%, while the same turns without feedback stay at 0.5–11.0%.
2️⃣ Designing the experiments matters. Replaying a system's own best probes gives it exactly the same evidence, yet in AlienCode 9 of 10 systems do worse than when they chose the probes themselves.
3️⃣ Knowing a rule is not using it. Even when every required rule is stated correctly, tasks are solved only 73.4% of the time.
4️⃣ One score hides a lot. The same system under the same budget ended anywhere from 5.7% to 79.0%, and rankings barely transfer between the two worlds.
CL-bench asked whether models can learn from context. ExplorationBench asks whether they can discover the rules themselves.
📄 Paper:
🌐 Website & leaderboard:
📝 Blog:
💻 Code (coming soon):
Show more
> Hy-MT2: translate this post
> to honest english
"it's a decent translator,try it"
440 MB,33 languages and pocket-sized
take it,ship it
gracias for the signal boost
🚨 ACABAN DE FILTRAR UN TRADUCTOR CON IA QUE LE GANA A TODOS Y ES GRATIS
Lo hizo Tencent y funciona sin internet: pesa 440 MB y cabe en tu celular.
Es open source, cualquiera puede descargarlo, usarlo e integrarlo en sus apps.
Se llama Hy-MT2.
Lo que trae:
→ 33 idiomas, incluido el español
→ Traduce 100% en tu dispositivo, incluso en modo avión
→ Tus textos no salen de tu teléfono
→ Supera a la API de traducción de Microsoft en las pruebas de Tencent
Sin suscripción. Sin mandar tus datos a ningún servidor.
Te dejo el repo abajo 👇
Show more
Stop uploading files to AI. Open them with AI instead.
Dragging files into chat boxes is pure habit by now, and it always felt a little backwards. So we built the other direction into WorkBuddy.
Right-click any document on your computer and open it with WorkBuddy: your AI is right there in the window, ready for questions. HTML, image, slides, Excel, Word and Markdown are editable too, and every change saves straight back to the original.
Files shouldn't have to travel to reach AI, and neither should you.
We're early on this road. Tell us where it should go next, and we'll build it with you.
Show more
The four core components behind Octop are now open source too: Octop Harness, Octop Memory, Octop Browser and Octop Gateway.
They handle agent execution, persistent memory, browser automation and messaging integrations. If you're building an agent and need one of these pieces, you can take it without the rest of the stack.
All four are MIT licensed, on GitHub and PyPI:
Show more
We open sourced Octop, a multi-user AI workspace for yourself, your family and your team.
Everyone gets their own login and can create multiple agents, each with its own workspace. Nothing you say or save shows up in another account. Almost everything can be swapped: models, storage, memory, plugins. Each agent comes with a browser, a terminal and a remote desktop.
You can hand off work to Claude Code, Codex, Cursor, OpenCode, CodeBuddy and other coding agents you already use. It also lives in your group chats on Telegram and Discord.
It runs on your own machine, and one command is all it takes: octop run.
The full picture is in the repo:
Show more
We open sourced Octop, a multi-user AI workspace for yourself, your family and your team.
Everyone gets their own login and can create multiple agents, each with its own workspace. Nothing you say or save shows up in another account. Almost everything can be swapped: models, storage, memory, plugins. Each agent comes with a browser, a terminal and a remote desktop.
You can hand off work to Claude Code, Codex, Cursor, OpenCode, CodeBuddy and other coding agents you already use. It also lives in your group chats on Telegram and Discord.
It runs on your own machine, and one command is all it takes: octop run.
The full picture is in the repo:
Show more
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.
Show more
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. 🕵️🐧
Show more
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:
Previews exist to get corrected, tell us what's off and it goes into the next version.
Show more
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:
Show more
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? 👀
Show more
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
Show more
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:
Show more
