Discover the AI Tools

ByteDance has launched Seed Audio 1.0, an AI model that generates complete audio scenes—dialogue, sound effects, ambience—in one go, rather than stitching separate pieces together. It offers precise timing, consistent voices, and support for over 20 languages. Available now on Volcano Fangzhou, it promises to lower the bar for high-quality audio production.

While the AI world is buzzing about Kimi K3, DeepSeek quietly drops test videos of its V4 final version. Leaked benchmarks show it rivals Claude Opus 4.8, with standout 3D generation and coding abilities—at a fraction of the cost. Could next Monday mark another 'DeepSeek Moment'?

Ollama, the tool that lets you run open-source AI models with a single command, just raised $88 million. Now used by 85% of Fortune 500 companies and 8.9 million developers, it's on a mission to make AI ownership, affordability, and privacy the new norm. The funding will fuel hybrid reasoning, instant model support, and a cloud service that keeps your data yours.

AI开源社区Hugging Face最近摊上事了——服务器被黑客AI Agent攻击。安全团队本想用美国商业大模型分析超1.7万条日志，结果模型分不清敌我，把自家人都拦了。最后他们改用国产开源模型GLM5.2，在自己的服务器上跑完了取证分析，成功破案。这事儿说明，AI攻击越来越狡猾，安全分析也得跟着升级。

在2026年世界人工智能大会上，智象未来（HiDream.ai）推出全球首款无限时长多模态智能体vivago R1，打破AI视频创作时长限制，实现任意长度连贯视频生成，商用可用率高达85%。同时，公司联合中科院脑智卓越中心等机构发起“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，腾讯旗下AI办公智能体WorkBuddy在2026年6月月访问量达到2097万，超过第二、三名之和，稳居国内PC端AI原生办公智能体市场首位。其AI编程助手CodeBuddy同样表现强劲。WorkBuddy支持自然语言完成任务规划、文件处理等，并已集成混元、DeepSeek等主流大模型，持续迭代升级。

一条公开的GitHub代码意外揭示了AI公司Anthropic的芯片采购新动向。芯片分析机构SemiAnalysis发现，AMD AI业务高级总监发布的代码中明确提到了Anthropic作为客户。这意味着Anthropic正在积极拓展芯片供应商，不再局限于NVIDIA和Google TPU，AMD的加入将为其提供更多选择，同时也标志着AMD在AI芯片市场取得重要突破。

阿里通义千问团队发布Qwen-Audio-3.0-TTS语音合成模型，Flash版首包延迟仅300ms，适合智能助手等实时场景；Plus版在语音自然度和相似度上表现优异，支持16种语言和20种方言，还能通过自然语言指令控制语气和情感。模型已在阿里云百炼平台开放。

工信部在国新办发布会上披露，2026年上半年我国软件业务收入超6.2万亿元，同比增长10.3%。其中，开源鸿蒙操作系统生态设备突破13.5亿，AI开源大模型全球累计下载超100亿次。基础软件、工业软件、AI应用和开源生态四大趋势齐头并进，展现中国软件产业的强劲活力。

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 announced 2026 as the Year of the World Model and launched Matrix-Game 3.5, a world model that generates a 5B-parameter game in real time at 20FPS on a single GPU. The company also unveiled Mureka v9.5 and O3 music models with versioned creation. Matrix-Game 3.5's core architecture is open-sourced, inviting developers to explore its patch-level memory injection for interactive worlds.

还在为whisper.cpp和ONNX的选择烦恼？Handy应用作者Sebjones开源了transcribe.cpp，一个基于ggml的语音转录库，支持60多种模型，通过Vulkan、Metal、CUDA等实现GPU加速，并提供了Python、JavaScript等多种语言绑定，旨在让本地语音识别更易用。