InclusionAI

✅We're also open-sourcing SingGuard — a policy-adaptive multimodal guardrail model family for image and text generation. SingGuard is designed to help developers build AI applications that better protect user privacy and intellectual property while reducing harmful or illegal outputs across multimodal scenarios. By making these models openly available, we hope to advance transparent evaluation, collaborative research, and safer AI development across the community. 🔗Paper: arxiv.org/abs/2606.22873 🔗GitHub: github.com/inclusionAI/Si… 🔗HuggingFace: huggingface.co/collections/in… 🔗ModelScope: modelscope.cn/collections/in… #OpenSource #AI #AISafety #Multimodal

Today we're open-sourcing SingGuard-NSFA, a security guardrail framework purpose-built for autonomous AI agents. As AI evolves from generating content to taking actions, security must move beyond model safety to runtime safety. SingGuard-NSFA helps defend against agent-specific threats such as prompt injection, goal hijacking, tool misuse, and permission abuse before actions are executed. The release includes: 🔐A taxonomy covering 185 agent attack scenarios across 7 categories 📊 A multilingual benchmark with nearly 100K test samples in 133 languages 🔧 Lightweight models optimized for real-world deployment We hope this project helps researchers and developers build safer, more trustworthy agentic AI. 🔗GitHub: github.com/inclusionAI/Si… 🔗Hugging Face : huggingface.co/collections/in… 🔗ModelScope: modelscope.cn/collections/in… #inclusionAI #OpenSource #AI #AISecurity #AgenticAI

Today we introduce LingBot-VA 2.0, the 🎆finale and also the capstone of LingBot launch week. 1️⃣ week, 6️⃣ models. Together they form our Embodied-Native Full-Stack — perception, world simulation, prediction, and action generation in one closed loop. One brain, many robots. 🔗 Project Page: technology.robbyant.com/lingbot-vla-v2 🔗Paper: github.com/Robbyant/lingb… #inclusionAI #LingBot #EmbodiedAI

What a night at ICML 2026! Tonight, we joined forces with the @sgl_project @radixark community to host a dinner for students, bringing together great food, fresh ideas, and even better conversations. The evening featured talks on the @AntLingAGI model series, SGLang/Miles, and LLaDA. A big thank-you to everyone who joined us, and to @ZenMuxAI for sponsoring the tokens that powered the night! See you around ICML! #ICML2026 #Ling #SGLang #Miles #LLaDA #OpenSourceAI

🤔Robots need physical grounding, not just visual beauty. Built for embodied AI — 🔥LingBot-Video , the first MoE-based video foundation model built for embodied intelligence. LingBot-Video is now fully open-source. 🔗 Model: huggingface.co/collections/ro… 🔗 Code: github.com/Robbyant/lingb… 🔗 Paper: github.com/Robbyant/lingb… 🔗 Project Page: technology.robbyant.com/lingbot-video

Meet LingBot-World 2.0 (Infinity) 🔥 The first interactive world model with: ✅ Hour-long generation with zero quality drift ✅ Rich actions & events — attack, cast spells, shoot, summon storms ✅ Agentic world — a Director Agent drives real-time world evolution Try it online: reactor.inc/lingbot-world-… 💻 Github: github.com/Robbyant/lingb… 🤖Hugging Face: huggingface.co/collections/ro… Website: technology.robbyant.com/lingbot-world-…

Live at @icmlconf in Seoul! Join Ant Group for a cup of coffee, explore our latest research through our posters, and chat with Tech Leaders from inclusionAI. We’d love to talk about everything around the @AntLingAGI Ling Model series, LLaDA, AI safety, reasoning, and post-training. Come say hi! #ICML2026 #AGI #llm

Powered by the new LingBot Vision 👀🤖 huggingface.co/collections/ro… - Apache 2.0 - 4 versions: small/base/large/giant 🔥 - Pretrained with masked boundary modeling

LingBot brings a 6B vision-language-action model to ModelScope, built for practical real-world manipulation across different robot bodies. 🤖 License: Apache 2.0 🔗modelscope.ai/models/Robbyan… 📄 modelscope.ai/papers/2508.02… 🏆 On tabletop bimanual manipulation, LingBot-VLA 2.0 reaches 66.2 progress / 34.4 success on Agilex Cobot Magic, ahead of LingBot-VLA 1.0, pi0.5, and GR00T N1.7. 🦾 Trained with 50,000 hours of robot trajectories across 20 robot configurations, plus 10,000 hours of egocentric human videos. ⚙️ Supports arms, end-effectors, grippers, dexterous hands, waist, head, and mobile-base signals in one unified action representation. 🧠 Uses sparse MoE action experts for cross-embodiment scaling, plus dual-query distillation from LingBot-Depth and DINO-Video for current geometry and future dynamics.

Meet LingBot-VLA 2.0! 🔥Fully open-source with outstanding performance: 🦾 AgileX Cobot Magic (Progress Score / Success Rate): 66.2 / 34.4% 🦾 Galaxea R1 Pro: 34.6 / 15.6% · GitHub: github.com/Robbyant/lingb… · HuggingFace:· huggingface.co/collections/ro… · Tech Report: github.com/Robbyant/lingb… · Website @robbyant_brain: technology.robbyant.com/lingbot-vla-v2 #inclusionAI #LingBot #EmbodiedAI #Robotics #OpenSource #VLA

New release today🚀Introducing LingBot-Depth 2.0, a major upgrade for robot's "eyes". Powered by LingBot-Vision — the 1️⃣first visual foundation model to use "boundary structure" as a pre-training objective. 🤗huggingface.co/collections/ro… 📷github.com/Robbyant/lingb… #inclusionAI #LingBot #Robotics #DepthEstimation #OpenSource #EmbodiedAI

Fully Autonomous Operation. What's driving them? The answer starts tomorrow. 🔥 Stay tuned with @robbyant_brain! #LingBot2 #inclusionAI #Robotics #EmbodiedAI

🙌Welcome to the AReaL community. First tweet.💥Big release. We are excited to announce the release of AReaL 2.0: RL as Micro-Service! Thanks to the great efforts from the AReaL community and collaborators❤️ In AReaL 2.0, we rebuilt online agent RL around four decoupled components: Training Service, Inference Serivce, WeightUpdate Service and Agent Service. Change one URL in your existing agent, start learning from production traffic. #AIAgent #agenticRL #AReaL #LLM 💻 GitHub: github.com/areal-project/…

AReaL 2.0 provides a service-oriented stack for agent RL: ✅ Gateway as the entry point ✅ Router for session affinity ✅ Data Proxy for trajectory capture ✅ Agent-Compute Worker for inference/training backends ✅ Online weight updates The release includes ecosystem updates and practical agent RL examples: Hermes Agent online RL, Claude Code Agent RL infra, LoRA RL workflows with @Macaron0fficial etc. Join the AReaL community in building open infrastructure for self-evolving agents. 🤖 GitHub: github.com/areal-project/… 📰 Tech Report: github.com/areal-project/…

📣 Introducing AReaL 2.0: an online RL architecture upgrade for agent applications with a core feature of RL as Micro-Service. 💡A step toward agents that can learn from real deployment experience. The @AreaL_AI team rebuilt Agentic RL infrastructure around decoupled services: Training Service, Inference Serivce, WeightUpdate Service and Agent Service. Plug your existing agent in by changing one URL.

New paper release from @AntLingAGI ➡️Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe 🧠 Key insight: FP4 training quality is not only about bit width, but also grid geometry. Our takeaway is simple: E2M1 should remain useful for range-limited workloads, but future FP4 training systems should support uniform 4-bit grids as first-class primitives. More details at arxiv.org/abs/2606.20381

🤔Real-world financial docs are messy. Camera blur. Faded ink. Tables spanning 10+ pages. 🔥Introducing FinixDoc, a new agentic parser trained to handle exactly this chaos: → Homoglyph-aware contrastive learning for visual robustness → Multi-stage RL reward shaping (text · layout · tables · reading order) → Human-in-the-loop data factory for continuous improvement 📊Results on FinixDocBench: 81.43 overall score, 92% parsing success on ultra-long pages 🤗Dataset: huggingface.co/datasets/inclu… 📄Paper: openreview.net/forum?id=kjgpN… #DocumentAI #OCR #FinTech #OpenSource