Discover the AI Tools

Moonshot AI announced that Kimi K2.5, its first trillion-parameter multimodal model, will retire at the end of this month. Launched and open-sourced in January, K2.5 was a versatile model supporting visual and text inputs, thinking and non-thinking modes, and dialogue and Agent tasks. Its successor, Kimi K3, open-sourced in July, boasts 2.8 trillion parameters and a 1-million-token context window, marking a new milestone in open-source AI.

Alibaba DAMO Academy, with Shengjing Hospital, developed DAMO LiON, an AI model that detects small liver lesions on CT scans. In a two-month trial, it found 15 overlooked malignant tumors, mostly around 1 cm. The AI also cut reading time by 27% and boosted sensitivity by 11.5%, helping junior doctors perform like seniors. Published in Nature Medicine, this tool acts as a safety net, catching what human eyes might miss.

Luke Metz, a prominent AI researcher, has left OpenAI to join Meta's Superintelligence Labs, marking another high-profile move in the intensifying competition for top talent. Metz's career has been a revolving door between leading labs, and his latest switch underscores how crucial researcher loyalty has become in the race toward superintelligence.

Luke Metz, a prominent AI researcher, has left OpenAI to join Meta's Superintelligence Labs, marking another high-profile move in the industry's talent war. Metz's career has been a whirlwind, with stints at OpenAI, Thinking Machines, and back again. Now he's part of Alexandr Wang's team at Meta, which has been aggressively recruiting top minds. This move underscores Meta's commitment to leading in next-gen AI.

8月24日，阿里正式推出视频生成大模型万相3.0，单次生成时长提升至30秒，并首次支持doc、xls、ppt、pdf、md等多种文档格式直接转视频。该模型在人物、道具、空间关系及视觉风格上保持高度一致，力求还原真实世界的物理与逻辑规律，为视频创作带来全新可能。

Alibaba's DAMO Academy has developed an AI model that can spot tiny liver tumors often missed by human eyes. In a two-month trial, it helped doctors find 15 malignant growths in over 10,000 patients, many just a centimeter wide. The model, published in Nature Medicine, also cut reading time by 27% and boosted detection accuracy, even helping junior doctors perform like seniors.

At WAIC 2026, the spotlight shifted from raw model power to real-world value. With 83 out of 130 exhibitors focusing on AI agents, the industry is moving beyond single models toward integrated systems. This article explores how AGI is becoming an organizational goal, why agents still depend on pre-trained models, and how AI-native programming is reshaping software development—offering a glimpse into the dawn of artificial general intelligence.

Shaoyn has launched the OpenFit 2 AI earphones, integrating the Qianwen large model for advanced voice features. Priced at 1398 yuan, these open-ear headphones offer 48-hour total battery life, IP55 protection, and comfortable design. With AI recording, translation, and meeting summaries, they redefine smart audio.

据美国商业内幕网报道，全球最大的AI开源社区Hugging Face正考虑整体出售，潜在交易估值可能达到130亿美元甚至更高。这家被称为“AI界GitHub”的公司，目前正与银行合作评估潜在买家的兴趣，但尚未达成任何协议。若交易成行，其估值将较2023年的45亿美元翻近三倍，也反映出资本对AI基础设施和分发平台的兴趣日益浓厚。

小米正式推出首款面向边缘大模型的高带宽AI加速芯片——玄戒O100。该芯片采用全球首创的6nm晶圆级垂直堆叠封装和混合键合技术，键合间距仅1.4微米，带宽高达1.22TB/s，是传统旗舰手机的16倍，边缘推理速度可达330TPS。这标志着手机AI竞争正从单纯算力转向存储带宽、封装技术和边缘大模型的综合优化。

Yahoo is betting on its vintage brand and a new AI-powered answer engine called Scout to win over Gen Z. CEO Jim Lanzone believes decades of proprietary data will set it apart from competitors. With Scout launching later this year, Yahoo aims to prove it's still relevant in the AI era, even as it lags behind tech giants.

小米正式推出首款专为大模型设计的AI加速芯片——玄戒O100。该芯片采用先进的6nm工艺和3D晶圆级堆叠技术，内存带宽高达1.22TB/s，运行小米MiMo 3B模型时推理速度可达330 Tokens/s。这标志着小米在自研核心计算芯片上迈出了关键一步，为终端AI应用提供了坚实的硬件基础。

South Korea's parliament has passed a landmark amendment to its Personal Information Protection Act, allowing AI developers to use limited personal data after regulatory review. This move aims to boost the local AI industry while maintaining privacy safeguards. The new rules, effective six months after State Council approval, replace the previous strict consent requirements with a review process, balancing innovation and data protection.

ByteDance is consolidating its AI office products, TRAE and Kousi, under the Doubao umbrella. This move aims to unify resources and strengthen Doubao's position in the competitive AI office market. The company plans to launch a dedicated 'Doubao Work' product soon, signaling a shift towards integrated AI work platforms.

American legal AI unicorn Harvey has developed a legal-specific model, Harvey Tenet, based on Moonshot AI's open-weight Kimi K3. This marks the first time a Chinese open-weight model has entered the training system of a leading US vertical AI company. Harvey, valued at $11 billion, previously relied on closed-source models like OpenAI and Anthropic. The move signals growing trust in open-weight models for specialized markets.

Xiaopeng Robotics just closed a massive first funding round, pulling in over $900 million and hitting a valuation of $6.3 billion. That's a record for China's embodied intelligence sector. IDG Capital led the round, with Tencent and Alibaba joining as strategic investors. The cash will speed up mass production of humanoid robots and AI model development, signaling a big shift from lab experiments to real-world deployment.

Oak Deer Robotics has released the world's first cooking AI foundation model trained on real kitchen data, marking a shift from simple commands to true understanding. Alongside it, two new robots—a visual stir-frying bot and an embodied cooking robot—promise to bring intelligent cooking into the real world. This could change how we think about automated meal prep.

OpenAI's Sam Altman recently voiced a concern that's been nagging at him: what if AI ends up controlled by a handful of powerful entities, leaving the rest of us without a say? In a podcast, he argued for a future where society and AI evolve together, with people having real control. He also took a jab at the 'open weights' debate, suggesting that calls for security might actually pave the way for monopolies.

OpenAI is addressing user complaints about rapid usage consumption in Codex. The company will reset quotas for all subscribers tomorrow and has identified three technical causes, including performance issues with images and context compression. Fixes are rolling out, with an unexpected efficiency improvement also in the works.

A new report from Guidelight AI Standards reveals that top AI labs are lagging in publicly disclosing plans to contain rogue AI. OpenAI scored highest with 3 out of 5 points, while Anthropic and Meta scored lowest. The assessment highlights the need for robust emergency shutdown mechanisms as AI agents become more autonomous, with new regulations on the horizon.