Jiesong Zhang, a member of Disheng Xu’s team and the paper’s author, developed ideas and arguments through dialogue with an internal version of Intern-S2. Bin Dong and Ruochuan Liu’s team at Peking University used SHEAF to formalise all four main theorems. Paper: arxiv.org/abs/2609.37695 Code: github.com/frenzymath/Holomo…

😃“From research questions to verifiable results” is the promise. 😉The proof? It ships. 🚀Meet Atria Dawn Preview @AtriaASI — an agentic foundation model for long-horizon tasks and controlled self-evolution, built on the Intern large model. 🌦️ 1 min → global weather forecast 🖥️ 20 min → OS from scratch 🧠 AI building AI, with verification in the loop Not uncontrolled RSI. Verifiable. Reproducible. Controllable. Open preview ↓ 🌐 atria-asi.ai 💻 github.com/atria-asi/Atria-D… 🤗 huggingface.co/internlm/Atri… 🤖 modelscope.cn/models/Shangha… 🐦 xitter.cf/AtriaASI #AtriaDawn #AgenticAI #RSI #OpenSource

🚀 Introducing Intern-S2-397B, our new multimodal foundation model designed for long-horizon scientific research and powerful scientific agents. 1⃣Strong general capabilities across knowledge, coding, and agents, ranking among the top open-source models. 2⃣Outstanding performance on scientific tasks including Biology-Instructions and Mol-Instructions, with strong capabilities in scientific understanding, generation, and design. 3⃣Achieves leading results among open-source models on IMO-Proof and AdvancedMathBench, reaching the level of top closed-source models such as Gemini 3.1 Pro. 4⃣Explores a new architecture for continuously expanding specialized scientific knowledge, building a stronger foundation for long-cycle scientific research. 👏Now supported by vLLM (@vllm_project ) and SGLang ( @lmsysorg ) — with more ecosystem integrations on the way. 🤗 Model: @huggingface huggingface.co/internlm/Inte… @ModelScope2022 modelscope.cn/models/Shangha… 🤗 Try it now: chat.intern-ai.org.cn

👏Proud to congratulate Prof. Bowen Zhou, Director and Chief Scientist of Shanghai AI Lab, on being named to the 2026 #TIME100AI “Thinkers” list. 🤗His vision inspires our work at Shanghai AI lab: advancing models for scientific discovery while strengthening AI safety and reliability. 😉Read more:time.com/collection/time100-…

Bowen Zhou, Director and Chief Scientist of Shanghai AI Laboratory，asked Nicholas B. Dirks whether external evaluations, red teaming, and third-party verification were enough to justify handing real-world authority to AI. "No," Nitzberg replied. He compared AI systems to bridges. Engineers do not build a bridge first and then attack it to discover whether it is safe. The builder carries the burden of proof during design. Increasingly powerful agents, Nitzberg argued, should likewise be safe by design, with evidence before deployment that their overall behavior remains understandable and controllable. That is why AI safety has become urgent now. Rules in aviation, finance, and medicine have often followed air crashes, financial crises, or public-health disasters. But AI cuts across all of those sectors. Its speed, scale, and cross-border reach make disaster an unacceptable starting point for governance. Nicholas B. Dirks pointed to a second difficulty: People are already using systems they do not trust, and adoption will continue to grow. Engineering verifiability does not automatically become social trustworthiness. Trust also requires independent oversight, clear responsibility, continuous monitoring, and assessment at every stage of deployment. #WAIC2026

The word "agent" carries two very different meanings. In AI, an agent is a system that can perceive, plan, and act. In law, an agent receives authority but also assumes a duty of loyalty and can be held accountable. Today's AI agents can receive the first of those things. They cannot assume the rest in the way a person or company can. Authority is beginning to flow to machines while responsibility remains in human society. Mark Nitzberg, the Executive Director of the Center for Human-Compatible AI and Head of Strategic Outreach at Berkeley AI Research，described this as a gap between engineering and law. A human agent who fails in a duty can be investigated and sanctioned. A probabilistic black-box system has no engineering equivalent of a fiduciary duty. Closing that gap, he argued, requires AI that is well-founded, legible, and steerable: built from components people can reason about, understandable in operation, and stoppable or changeable at any time. Lan Xue，Dean of Schwarzman College and the Institute for AI International Governance at Tsinghua University，noted that classic principal-agent problems have not vanished. Information asymmetry, moral hazard, and misaligned objectives all persist. What has changed is the accountability chain. A machine cannot be punished, so responsibility must be redistributed across design, development, deployment, and use. Nicholas B. Dirks, President and CEO of The New York Academy of Sciences，warned that delegation is never only about handing off a task. In the pursuit of efficiency, people may also delegate goals, judgment, and principles. "If we are only thinking about AI that way," he said, "we are already beginning to forget about our values." #WAIC2026 #agent

🔥Introducing Visual Pretraining (VP) 👀: a new, scalable pretraining paradigm for foundation models. 😉It enables models to learn directly from visual documents 📖 and outperforms text-only pretraining across backbones and benchmarks. 📈 Explore the paper and our most capable multimodal foundation model trained with Visual Pretraining as part of its training recipe. 📄 Paper: arxiv.org/abs/2607.09657 🤗 Hugging Face Paper: huggingface.co/papers/2607.0… 🤖 Intern-S2-Preview (35B): huggingface.co/internlm/Inte… 🤖 Intern-S2-Preview-397B: huggingface.co/internlm/Inte…

🚀Introducing Intern-S2-Preview-397B, our most capable multimodal foundation model for scientific intelligence and long-horizon agents.🔥 1⃣Delivers a step change in general reasoning, scientific problem solving, and agentic capabilities. 2⃣A new pre-training paradigm preserves text-visual correspondence, strengthens spatial and visual reasoning, and improves data efficiency. 3⃣Achieves leading general-reasoning performance among open-source models and strong results in specialized scientific tasks by scaling diverse scientific RL tasks across 20+ domains and training them jointly. 4⃣Improves generalization and raises the capability ceiling for long-horizon tasks in both general and scientific domains by connecting multiple agent frameworks to large-scale sandboxed environments for black-box agentic reinforcement learning. 😉Model： @huggingface huggingface.co/internlm/Inte… @ModelScope2022 modelscope.cn/models/Shangha… 😉GitHub: github.com/InternLM/Intern-S… 😉Try it now at: chat.intern-ai.org.cn

🥳Introducing Intern-S2-Preview, an efficient 35B scientific multimodal foundation model. 1⃣Delivers performance comparable to the trillion-scale Intern-S1-Pro on core scientific tasks. 2⃣The first open-source model with material crystal structure generation capabilities and strong general capabilities. 3⃣Significantly stronger scientific agent capabilities on multiple benchmarks. 4⃣Improves MTP acceptance rate and token generation speed via shared-weight MTP + KL loss. 5⃣CoT compression shortens responses while preserving strong reasoning , improving both performance and efficiency. 🥰Now supported by vLLM (@vllm_project) and SGLang ( @lmsysorg ) — with more ecosystem integrations on the way. 🤗Model： @huggingface huggingface.co/collections/i… @ModelScope2022 modelscope.cn/collections/Sh… 🤗GitHub: github.com/InternLM/Intern-S… 🤗Try it now at: chat.intern-ai.org.cn

🔥Introducing Kernel-Smith, a framework for high-performance GPU kernel and operator generation. 1⃣Combines a stable evaluation-driven evolutionary agent with an evolution-oriented post-training recipe. 2⃣Outperforms Gemini-3.0-pro & Claude-4.6-opus on Kernel-Bench. 3⃣Optimized kernels are already merged into SGLang @lmsysorg and LMDeploy. 😉Tech report: huggingface.co/papers/2603.2… 😉Try it now at: chat.intern-ai.org.cn/kernel…

🔥Introducing ARM-Thinker, the first Agentic multimodal Reward Model that autonomously invokes external tools to ground judgments in verifiable evidence. Accepted to CVPR 2026! 🥳Integrates three families of multimodal tools: 1⃣Image Crop & Zoom-in for fine-grained visual inspection. 2⃣Document Retrieval for multi-page evidence gathering. 3⃣Instruction-Following Validators for constraint verification. 🥳With a Think-Act-Verify loop, ARM-Thinker can call image crop & zoom-in, document retrieval, and instruction-following validators for evidence-based evaluation. 🥳Built on Qwen2.5-VL-7B with SFT + two-stage GRPO, ARM-Thinker improves multimodal reward modeling, tool-use reasoning, and multimodal math/logical reasoning. 😉+16.2% on reward modeling benchmarks (outperforming GPT-4o). 😉+9.6% on tool-use / think-with-images tasks (matching Mini-o3). 😉+4.2% on multimodal math & logical reasoning. 🥳Also introduce ARMBench-VL, the first multimodal reward benchmark that requires tool use. 📄 Paper: arxiv.org/abs/2512.05111 💻 Code: github.com/InternLM/ARM-Thin… 🤗 Dataset: @huggingface huggingface.co/datasets/inte… 🧪 Evaluation: github.com/open-compass/VLME…

🚀Meet InternVL-U: a lightweight 4B unified multimodal model that brings reasoning, generation, and editing into a unified framework. 🔥Built upon unified contextual modeling, modality-specific modular design, and decoupled visual representations, InternVL-U achieves a strong performance-efficiency trade-off, consistently outperforming unified baselines with over 3× larger model scales on challenging tasks such as text rendering, scientific reasoning, and spatially grounded generation and editing. 😉Open-source and designed for efficient, practical multimodal intelligence. 🤗GitHub: github.com/OpenGVLab/InternV… 🤗Hugging Face: @huggingface huggingface.co/InternVL-U/In… 🤗GenEditEvalKit: github.com/open-compass/GenE… 🤗TextEdit: github.com/open-compass/Text… 🤗Tech report: arxiv.org/pdf/2603.09877

🚀Introducing Intern-S1-Pro, an advanced 1T MoE open-source multimodal scientific reasoning model. 1⃣SOTA scientific reasoning, competitive with leading closed-source models across AI4Science tasks. 2⃣Top-tier performance on advanced reasoning benchmarks, strong general multimodal performance on various benchmarks. 3⃣1T-A22B MoE training efficiency with STE routing (dense gradient for router training) and grouped routing for stable convergence and balanced expert parallelism. 4⃣Fourier Position Encoding (FoPE) + upgraded time-series modeling for better physical signal representation; supports long, heterogeneous time-series (10^0–10^6 points). 😍Intern-S1-Pro is now supported by vLLM @vllm_project and SGLang @sgl_project @lmsysorg — more ecosystem integrations are on the way. ☺️Model：@huggingface huggingface.co/internlm/Inte… ☺️GitHub: github.com/InternLM/Intern-S… ☺️Try it now at: chat.intern-ai.org.cn

🚀 Introducing Spatial-SSRL, the first study which proposes a Self-Supervised Reinforcement Learning paradigm for spatial understanding. 💡 Spatial-SSRL a lightweight tool-free framework that aims at enhancing spatial intelligence and is natually compatible with the RLVR training paradigm. Only raw 2D and RGB-D images are required and we avoid any use of human annotation, external proprietary model or expert model throughout the entire pipeline, making Spatial-SSRL highly cost-effective and scalable. 🛰️ Spatial-SSRL comprises five pretext tasks now: shuffled patch reordering, flipped patch recognition, cropped patch inpainting, regional depth ordering, and 3D relative position prediction. Thanks to its lightweight characteristics, Spatial-SSRL can be easily extended to more pretext tasks and we welcome the whole community to join Spatial-SSRL with effective pretext tasks! 🤖 After applying Spatial-SSRL, we significantly enhance the performance of spatial understanding on Qwen2.5-VL (3B&7B) and Qwen3-VL (4B), as well as retaining their general visual capabilities. 🤗 Currently, we have released the repository of Spatial-SSRL, the dataset Spatial-SSRL-81k, and the trained models: Spatial-SSRL-7B and Spatial-SSRL-Qwen3VL-4B. The total download of the models and dataset has surpassed 1,000. 👇 Try Spatial-SSRL-7B now at: huggingface.co/spaces/yuhang… Paper: arxiv.org/abs/2510.27606 Github: github.com/InternLM/Spatial-… Model (on Qwen2.5-VL): huggingface.co/internlm/Spat… Model (on Qwen3-VL): huggingface.co/internlm/Spat… Dataset: huggingface.co/datasets/inte…

🚀Introducing #CapRL, the first study of applying GRPO for the open-ended and subjective image captioning task. 🤯 🤖The trained CapRL-3B model achieves image captioning performance comparable to Qwen2.5-VL-72B. ✨CapRL introduces a novel training framework that redefines caption quality through its utility: a high-quality caption should enable a non-visual language model to accurately answer questions about the corresponding image. 📈Currently, CapRL is open-sourced, with total downloads of the models and datasets surpassing 7,000. The research team is continuously iterating with stronger base models and improved training recipe. 👇 Try it now at: huggingface.co/spaces/yuhang… Paper: arxiv.org/abs/2509.22647 GitHub: github.com/InternLM/CapRL Model: huggingface.co/internlm/CapR… Dataset: huggingface.co/datasets/inte…

🚀 Big news for #lmdeploy v0.10.1! 🥳Our #FP8 high-performance inference is no longer limited to the latest #GPUs. It now supports all #NVIDIA architectures from V100 onwards, bringing major speedups to more users. 🤗github.com/InternLM/lmdeploy

🔥LMDeploy v0.10.0 released! 😊Supercharges OpenAI’s GPT-OSS MXFP4 models. 😊Delivers exceptional performance for GPT-OSS models on V100 and higher GPUs. 😊On H800 & A100, LMDeploy outperforms vLLM across all scenarios—faster, more efficient inference! 🤗github.com/InternLM/lmdeploy