Discover the AI Tools

Tencent's Hunyuan team has unveiled Hyra-1.0, a research agent that can recursively improve itself. It outperforms benchmarks in AI development, solves math problems, designs drugs, and even plays Go. The system follows a simple, lightweight design philosophy, turning computing power into real results. From optimizing GPU kernels to composing harmonies, Hyra shows promise across AI, science, and fun.

Microsoft is reportedly considering using a Chinese AI model, Kimi K3, in its Copilot service to slash costs by up to $600 million annually. But the move could trigger regulatory backlash from the US government, which is tightening restrictions on Chinese AI tech. The company faces a tough choice between huge savings and political risk.

在2026世界人工智能大会上，智元动力创始人张巍语出惊人：人形机器人正处在指数级增长曲线上，其大脑系统已相当于GPT-3阶段。他直言机器人比光刻机、飞机甚至汽车都好造，零件数只有汽车的十分之一。真正的瓶颈不在硬件，而在于物理世界的数据匮乏。智元动力已获阿里、京东等巨头投资，正加速商业化落地。

受全球AI和数据中心建设热潮推动，韩国7月前20天出口额同比猛增62.9%，创历史同期新高。其中半导体出口暴涨180.6%，电脑及周边产品更是飙升近232%，科技硬件成为出口主力。

China Telecom has handed over 5G network planning to a large AI model, achieving a 50% boost in efficiency and 75% accuracy in solutions. The model, tested in Shanghai, automates site selection and design, marking a shift from experience-based to data-driven planning. This move aligns with the industry's push for 'AI plus information and communication' and could redefine how telecom operators operate.

Neill Blomkamp, the director behind the acclaimed sci-fi film 'District 9,' has released his first AI-generated short film, 'Yin Bing.' The 13-minute sci-fi piece was created entirely using ByteDance's Seedance 2.0 model, with Blomkamp guiding each frame through text prompts. The film, which features 32 real actors and concept designers, explores a secret super-soldier program and signals a new era for AI in filmmaking.

Alibaba has released Qwen-Image-3.0, a new image generation model that supports up to 4.5K token text input. It can create complex visuals like formulas, diagrams, and multi-layered UI designs in one go. The model also supports 12 languages and over 20 fonts, making it a powerful tool for commercial content creation.

XPeng Group has unveiled TuringViT, an efficient visual encoder that achieves state-of-the-art performance using only one-tenth of the data required by competitors. By rethinking architecture, data curation, and training, TuringViT makes advanced vision AI more accessible for applications like autonomous driving and robotics.

月之暗面公司的新模型Kimi3上线后，用户需求远超预期，导致计算资源紧张。公司暂停新用户注册，优先保障付费用户体验。团队正通过优化模型效率和补充算力来缓解压力，恢复注册时间待定。这反映出AI行业竞争已从模型能力扩展到算力、基础设施和服务稳定性。

Alibaba is launching 'Qwen Office,' a new AI-powered office product that combines three of its intelligent assistants: QoderWork, Wukong, and MuleRun. Led by DingTalk's new CEO, Chen Yusen, the move aims to create a unified AI office entry point. QoderWork will serve as the core platform, while the integration signals Alibaba's push to compete in the rapidly growing AI office agent space, leveraging its cloud and large model ecosystem.

YouTube is cracking down on AI-generated spam. New rules ban three types of 'non-real' content from monetization: repetitive templated videos, offensive or distressing clips, and videos using virtual characters to discuss sensitive topics like health and finance. The goal is to boost content quality and encourage original work.

OpenAI宣布升级ChatGPT的家长通知功能，当未成年用户因暴力威胁或网络霸凌被禁时，关联账户的家长将收到警报。该功能与反霸凌组织Moonshot合作开发，旨在帮助家庭及早干预。此前OpenAI已推出家长控制功能，此次升级进一步覆盖可能伤害他人的风险行为。

A new analysis reveals that Anthropic's Fable 5 model has inference costs three times higher than OpenAI and open-source alternatives, raising questions about its market viability. As Chinese models like Kimi K3 and Qwen 3.8 match its performance at lower cost, the competitive landscape shifts. The analysis categorizes AI companies by infrastructure ownership, highlighting cost advantages for those with data centers or self-generated power.

Sakana AI's new multi-agent system, Fugu Cyber, has achieved top scores in cybersecurity benchmarks, outperforming OpenAI's GPT-5.5-Cyber and Anthropic's Claude. The model scored 86.9% on CyberGym and 72.1% on CTI-REALM, signaling a shift in how AI can tackle complex network defense challenges.

Tencent Smart Retail and Tencent Maps have teamed up with Chinagoods to launch an AI-powered navigation system in Yiwu International Trade City. Covering 80,000 shops across 8 million square meters, the solution offers indoor-outdoor navigation, AI product recognition, and multilingual support, cutting store-finding time by nearly 70%.

OpenAI前CTO Murati创办的Thinking Machines Lab发布首个模型Inkling，架构借鉴DeepSeek-V3，训练数据用了Kimi K2.5。结果多项评测落后于中国开源模型，价格却更高。但仔细想想，这或许不是技术失误，而是一场精准的商业算计：瞄准美国企业不敢用中国模型后的合规替代市场。

Embodied intelligence startup Zhiyuan Robotics is reportedly pushing forward with an IPO at a $20 billion valuation, having hired CITIC Securities as a sponsor. The company expects to hit 4 billion yuan in revenue for 2026. At the recent WAIC, it unveiled five new robot models and showcased its integrated AI and robotics ecosystem, signaling a push toward large-scale commercialization.

Google DeepMind unveils GenCeption, a single AI model that handles depth estimation, segmentation, 3D pose estimation, and more—all at once. Trained on just 7,500 synthetic videos of single people, it surprisingly excels on real-world multi-person scenes, animals, and robots. The model processes 81 frames in about 6 seconds, with detail so sharp it preserves cat whiskers and individual hairs.

宇树科技最近放了个大招——推出UnifoLM-OminiA-0.3通用智能模型。这玩意儿能让机器人听懂人话、看懂东西，还能自己琢磨着干家务、照顾老人。从搬垫子到整理衣服，甚至调病床高度，机器人G1都能搞定。看来，人形机器人离走进日常生活又近了一步。

A new report reveals that the five largest US tech companies have accumulated a staggering $1.65 trillion in hidden debt—eight times more than four years ago. Much of this comes from data center leases and GPU supply contracts, with Meta alone owing $420 billion off the books. This lack of transparency makes it tough for investors to gauge financial health, especially as the AI arms race fuels spending.