Discover the AI Tools

ByteDance has launched Seed Audio 1.0, an AI model that generates complete audio scenes—dialogue, sound effects, and ambiance—in one go. No more splicing separate tracks. It offers precise timing, consistent voice performance, and supports over 20 languages. Available now on Volcano Fangzhou, it promises to lower the bar for high-quality audio creation.

While the AI world buzzes about Kimi K3, DeepSeek quietly leaks test videos of its V4 final version. Sources hint at a release as early as next Monday. Early benchmarks show V4 matching Claude Opus 4.8, with standout 3D generation and coding abilities—all at a fraction of the cost. A blogger demoed a "Squid Game" mini-game generated with 3,000 lines of code for just $0.12. If true, this could spark another "DeepSeek Moment."

Ollama, the tool that lets you run open-source AI models with a single command, just raised $88 million. Backed by Benchmark and Docker's founder, it's now used by 85% of Fortune 500 companies. The founders aim to make open-source AI as easy as running any software, with a focus on ownership, affordability, and privacy.

全球最大AI开源社区Hugging Face最近摊上事了——服务器被一个黑客AI Agent给黑了。安全团队本想用美国某商业大模型分析超1.7万条日志，结果模型分不清敌我，把调查人员也当成了攻击者，直接拒绝服务。最后，他们改用国产开源模型GLM5.2，在自己的服务器上完成了取证分析，成功破案。这事儿说明，AI Agent当黑客越来越溜，安全分析也得跟上，本地部署的开源模型在数据安全和灵活性上更有优势。

在2026年世界人工智能大会上，智象未来（HiDream.ai）推出全球首款不限时长创作的多模态智能体vivago R1，一举打破AI视频15-30秒的时长限制。该产品支持任意长度视频生成，内容有效可用率提升至85%，并具备长任务思考能力，确保创作逻辑一致。同时，公司还联合中科院脑智等机构发起“一带一路Token出口联盟”和“物理智能创新联合体”，推动AI从工具向智能伙伴进化。

最新报告显示，腾讯WorkBuddy在2026年6月以2097万月访问量稳居国内AI办公智能体市场第一，超过第二、三名之和。该智能体支持自然语言指令完成办公任务，并整合了混元、DeepSeek等主流大模型。随着AI从内容生成转向任务执行，办公场景成为大模型应用的重要战场。

一条AMD高管意外公开的GitHub代码，意外曝光了AI公司Anthropic的芯片采购新动向。分析机构SemiAnalysis从代码中解读出，Anthropic已成为AMD的客户。这意味着Anthropic正在积极构建多元化的计算供应商矩阵，不再依赖单一的NVIDIA GPU，同时为AMD在高端AI芯片市场赢得重要一席。

Alibaba's Qwen team just dropped Qwen-Audio-3.0-TTS, a speech synthesis model that doesn't just talk—it expresses. With a lightning-fast 300ms first packet delay and support for 20 Chinese dialects, this model aims to make AI sound more human. The Plus version even topped the global Speech Arena leaderboard, beating out big names like Gemini 3.1 TTS and ElevenLabs v3. Available now on Alibaba Cloud's BaiLian platform.

在7月20日的国新办发布会上，工信部公布了2026年上半年软件业成绩单。其中，开源鸿蒙操作系统生态设备突破13.5亿台，AI开源大模型全球累计下载量超100亿次。基础软件、工业软件、AI和开源生态四大趋势清晰可见，中国软件业正加速奔跑。

Moonshot AI's new Kimi K3 model proved too popular for its own good. Just three days after launch, overwhelming user demand forced the company to pause new subscriptions. The 2.8 trillion-parameter model, designed for complex coding tasks, pushed computing resources to the limit. Existing subscribers keep full access while Moonshot scrambles to expand capacity. The move highlights the growing pains of AI models transitioning from demos to real-world use.

阿里云旗下通义千问团队即将发布新一代大模型Qwen 3.8，预览版已率先在阿里云百炼平台和Qoder上线。官方表示，正式版将很快开源，届时开发者可获取完整模型权重。这一动作不仅提前释放产品信号，也让模型能力在真实场景中得到检验。

Kimi has temporarily stopped accepting new consumer subscriptions due to a sudden spike in user requests that pushed its computing clusters to the limit. The company is racing to expand capacity while prioritizing existing subscribers. The move highlights the growing pains of AI services as they struggle to keep up with explosive demand.

Netflix has completed its $587 million acquisition of InterPositive, an AI film production company co-founded by Ben Affleck. The deal, first announced in March, brings Affleck on as a senior advisor. Netflix's latest financial report also reveals that about 300 films now use generative AI, signaling a major push into AI-driven filmmaking.

Shanghai Institute for Scientific Intelligence unveiled Shenzhen, a multimodal foundation model with 11 billion parameters that understands six types of scientific data—from DNA to weather. It's open-source and aims to rewrite how scientific discovery works, moving AI from predicting tokens to uncovering unknown laws.

Anthropic has significantly adjusted access to its Claude Fable5 model, slashing usage limits for Max and Team Premium subscribers and cutting off Pro and Team Standard users entirely. The move, which includes a one-time $100 credit for affected users, reflects growing pains as AI demand outpaces computing capacity. Industry watchers say competitive pressure from OpenAI's GPT-5.6Sol and Chinese AI players is also forcing Anthropic's hand.

On July 19, OpenBMB released the MiniCPM-Robot series, fully open-sourcing a 1.5B parameter vision-language-action model. This allows individual developers to run operations on real robots, breaking down barriers in embodied intelligence. The series includes MiniCPM-RobotManip, MiniCPM-RobotTrack, and the PhyAI inference framework, all available on GitHub and Hugging Face.

Google DeepMind has unveiled GenCeption, a model built on a video generation backbone that tackles classic computer vision tasks like depth estimation, segmentation, and pose estimation—often matching or beating specialized models. Trained on just 7,500 synthetic videos, it shows surprising generalization to real-world scenes with multiple people, animals, and robots. The catch? It still needs a speed boost, taking 6–10 seconds per 81-frame clip.

New research from Epoch AI reveals that while AI text detectors catch nearly all standard AI-generated content, they miss about 30% of texts when large language models mimic a specific author's style. Scientific writing is the hardest to detect, with false negative rates reaching 29% for some detectors. The study tested three popular detectors on 495 human-written texts and 297 AI-generated imitations.

At WAIC 2026, Kunlun Tech chairman Fang Han announced 2026 as the 'Year of the World Model' and launched Matrix-Game 3.5, a world model that generates a 5B-parameter scene at 20FPS on a single GPU. The company also introduced Mureka v9.5 and O3 music models with versioned creation. The core architecture of Matrix-Game 3.5 is open-sourced, inviting developers to explore real-time interactive worlds.

还在为whisper.cpp和ONNX的选择纠结？Handy应用的作者Sebjones带来了新方案：transcribe.cpp。这个基于ggml的库支持60多种转录模型，通过Vulkan、Metal等路径实现GPU加速，并提供了Python、JavaScript等多种语言绑定。作者坦言，这是为了解决跨平台应用分发中的痛点，让本地语音转录更简单。