Discover the AI Tools

豆包的任务模式迎来重大更新，新增技能、连接器和工作伙伴三大功能，标志着其从聊天助手向智能办公平台的转型。目前平台已提供超过200个技能和连接器，用户还能自定义技能，打造标准化工作流。这些功能可单独使用或自由组合，帮助用户简化复杂流程，提升办公效率。

SAP has joined forces with esports powerhouse Team Liquid to create an AI voice application that turns in-game chatter into structured data. This tool could change how teams analyze matches and train players, offering a new lens on communication and decision-making in the heat of competition.

SenseTime has open-sourced SenseNova U1.5Lite, an 8B-parameter multimodal model that punches above its weight. It handles 4K resolution natively, edits images with precision, and excels at text-heavy layouts. This lightweight model could be a game-changer for developers and creators who need powerful visual AI without the heavy compute.

Liquid AI and Hugging Face have released DSpark draft models for the LFM2.5 series, boosting inference speed by up to 3.18x on GPU and 2.87x on edge devices. The technology uses speculative decoding to cut latency, with function calls on edge devices improving by 57%. It's compatible with SGLang and llama.cpp, and maintains output quality.

Mistral AI has launched Agentic Search, a multi-step retrieval system that boosts accuracy on complex documents to 86%. Unlike traditional RAG, it uses five tools—search, open, navigate, read, and grep—to reason and iterate. In tests on financial and government documents, accuracy tripled with the search loop, and full toolchain reduced latency and token use. This innovation promises to unlock value in sensitive enterprise data.

Alibaba has open-sourced Qwen-UI-Agent, a multimodal foundation model designed to operate real devices—not just simulations. It excels in mobile, desktop, and web tasks, scoring high on benchmarks and handling real-world complexities. With built-in safety checks and efficient command execution, it's a significant step toward practical AI assistants.

OpenAI has open-sourced Codex Harness, the core runtime behind its Codex AI coding agent. This move gives developers access to the framework that manages tasks, context, and tool execution, enabling them to build custom AI programming tools. The release signals a shift in AI coding competition from model power to agent infrastructure, with implications for how software is developed.

Kakao's board has approved a split to create a new AI-focused company, Kakao AI, while the existing entity becomes Kakao X. The move aims to accelerate AI integration into KakaoTalk, turning it into an intelligent assistant. Shareholders will receive shares in both companies, with the split expected to complete by early 2027.

At its 2026 earnings briefing, iFLYTEK announced a new flagship general-purpose large model, built entirely on domestic computing power. With top-tier coding abilities and cost-efficiency, it's set to debut at the 1024 Developer Festival. A phased version arrives by end of August, promising a significant leap for China's AI self-reliance.

TaoTie has unveiled its iO AI glasses, priced at 2499 yuan (promo 1996 yuan), featuring a dual-eye display, 24/7 AI memory, and privacy-first design. The glasses support prescription lenses, translate 55 languages, and are available on major Chinese platforms. With a lightweight 34-gram frame and long battery life, they aim to be a daily companion. But is the AI memory feature a gimmick or a game-changer? Let's dive in.

DeepSeek has rolled out a surprise update to its Harness tool, introducing a new visual model experience version called V4Flash Vision-Exp. This update brings multimodal capabilities to the development toolchain, allowing users to send images along with text commands. The move completes DeepSeek's multimodal puzzle, making it easier for developers to integrate visual understanding into their workflows. With just an npm command, developers can now access this speed-and-accuracy combined model.

ByteDance's Seed division has undergone a major restructuring, creating four new departments to streamline AI development. This move integrates data, pre-training, and reinforcement learning, while separating enterprise and consumer product strategies. The reorganization aims to boost model intelligence and tailor AI for office and chat applications. Here's what this means for the future of AI.

Google DeepMind's open-source Gemma models have surpassed one billion downloads, spawning over 100,000 variations. The models are now powering space missions with NASA, aiding medical diagnostics in India and Uganda, and even decoding dolphin communication. This milestone highlights Gemma's versatility across extreme environments and its growing impact on real-world applications.

OpenAI has rolled out a significant update to ChatGPT's macOS app, enabling deep integration with Apple's Messages. With user permission, the AI can read, compose, and send texts, and even summarize chat histories. This move has caught the attention of Bloomberg's Mark Gurman, who sees it as a direct challenge to Siri's ecosystem advantages. The feature requires several permissions, including full disk access, and could signal a new era of AI-powered communication.

OpenAI has regained the lead in the U.S. enterprise AI market, surpassing Anthropic with over 40% share. Ramp's data shows shifting dynamics as spending on AI tools continues to climb. What's driving this change? New models, pricing strategies, and data policies are reshaping customer choices.

xAI旗下的聊天机器人Grok近日出现异常，部分用户在Grok.com上请求生成PDF后，收到了连续的无意义乱码文本，甚至包含强化学习研究相关的链接。问题主要影响Grok Lite用户，刷新或重启会话通常能恢复正常。官方回应称这是“罕见的临时生成故障”，并建议用户重启对话。此次事件发生在xAI人员流失严重的背景下，再次凸显了大模型在推理、生成和检索链路上的稳定性挑战。

Alibaba has unveiled Qwen-UI-Agent, a new GUI agent model that excels across mobile, desktop, and web environments. In head-to-head tests, it outperformed GPT-5.6Sol and Claude Opus4.8 on mobile benchmarks, scoring 82.1% on MobileWorld and 97.5% on AndroidDaily. The model also supports real-device training and robust security features, marking a significant leap in GUI automation.

在2026世界机器人大会上，宇树科技创始人王兴兴透露，公司正推进物理AI机器人模型的自我进化，并探索用大模型自动生成机器人控制代码。他认为，具身智能的“ChatGPT时刻”仍需突破泛化能力瓶颈，预计2到3年内可能迎来转折点。

Tencent's next-gen AI model, Hy4, has surfaced in internal tests, hinting at an imminent release. Positioned as an expert-level model, it's set to power the WorkBuddy assistant. With potential upgrades like trillion-scale parameters or advanced reinforcement learning, Hy4 aims to tackle complex tasks, building on the success of its predecessor Hy3.

Apple Music is making AI transparency labels mandatory for AI-generated tracks by late 2026, a shift from voluntary disclosure. The move is part of broader efforts to curb AI spam, including internal monitoring tools and royalty reallocation. With over a third of uploads now AI-generated but barely listened to, platforms like Deezer and Spotify are also tightening their rules, signaling a new era of accountability in music.