Today’s AI daily
可灵 AI 正式发布 Kling 4.0 视频生成模型，支持最长 30 秒一镜到底及多帧参考功能。
Original · translation pending可灵 AI 正式上线视频模型 Kling 4.0，支持最长 30 秒直出和最多 10帧的多帧参考，并强化全能参考与复杂打斗等大范围动态表现。1080P 30秒约 450 积分，官方还预告将支持 10-bit 4K HDR 输出。
Nous Research 的开源自主智能体 Hermes Agent 成功入选时代周刊 2026 年度最佳发明榜单。
Original · translation pendingNous Research 的开源自主智能体桌面应用 Hermes Agent 入选 TIME「The Best Inventions of 2026」榜单。该智能体可运行任意本地或云端模型，支持自定义界面、编写 skills、执行终端命令、编辑文件并自主安排周期性任务。
马斯克表示 xAI 在 2027 年底的算力目标可能超过 10GW，英伟达已将 SpaceX 列为云服务商客户并合作设计太空服务器。
Original · translation pendingMusk 转发黄仁勋在 BG2Pod 访谈中的表态并评论算力目标可能超过 10GW；访谈称 Nvidia 已把 SpaceX 归入云服务商客户，两家工程师正合作设计可用于太空的新型服务器架构。
Thinking Machines Lab 回顾后训练 API 平台 Tinker 一周年，用户可通过几行代码将领域专长微调进前沿模型。
Original · translation pendingThinking Machines Lab 发布后训练 API 平台 Tinker 上线一周年的回顾文章，并上线新官网和快速上手体验。Tinker 旨在让用户无需自建基础设施团队、采购 GPU 或调试训练流程，只需几行代码就能将领域专长微调进前沿模型，让普通人也能训练自有模型；据称其微调成本仅为 GPT-5.5 的十四分之一。
谷歌 Gemini 4 Argon 面向全部 Pro 订阅用户全量开放，并提供独立周配额。
Original · translation pending谷歌宣布 Gemini 4 Argon 在 Antigravity 编码环境面向全部付费 Pro 订阅用户全量推送，官方称其为迄今最强模型。Argon 配额单独按周计算，不挤占原有 3.8 Flash 的 Gemini 配额，界面显示周配额使用百分比与刷新倒计时。
OpenAI 推出 Codex Sites 新功能，支持接入定时任务自动刷新站点数据，并允许用户连接插件为每位访客呈现个性化视图。
Original · translation pendingOpenAI 官方发布 Codex Sites 新功能:团队站点可接入定时任务(scheduled tasks)自动刷新站点数据,保持内容最新;用户还能连接自己的插件,让同一站点为每位访客呈现个性化视图。官方称直接向 Codex 提出需求即可试用。
Musk publicly recommends the Spam-Call Scrubber skill created by a user based on Grok Bot, which can automatically delete phone numbers on doxxing and data broker websites.
Musk retweeted the @Grok Bot skill "Spam-Call Scrubber" created by a user. This AI bot can find users' phone numbers on doxxing websites and data broker websites, and submit opt-out or deletion requests on their behalf to help clean up leaked personal information on the internet, receiving a public recommendation from Musk.
SGLang 团队宣布对 Meta 816B 开源多模态 MoE 模型 Muse Spark Open 提供 Day-0 推理支持，提供 NVFP4 与 BF16 权重。
Original · translation pendingSGLang 团队宣布对 Meta 开源多模态 MoE 模型 Muse Spark Open（816B 参数、约 42B 激活、1M token 上下文）提供 Day-0 推理支持，权重提供 NVFP4 和 BF16 两种格式。
The Krea 3 image model is now available through ComfyUI Partner Nodes, featuring fast iteration driven by multiple reference images and low cost per generation.
The Krea 3 image model is now available in ComfyUI through Partner Nodes, focusing on reference image-based generation and editing. The model emphasizes faster iteration speed and lower cost per image, supporting referencing multiple reference images in the order they are uploaded, modifying environmental lighting, shadows, and composition while keeping product label shapes and materials unchanged, and allowing generated results to be used as new inputs for further editing.
Anthropic announced that Code Review is now officially available to Claude Team and Enterprise users, with review speed improved by about 3 times to 6-10 minutes.
Anthropic announced that Code Review is exiting research preview and officially launching to Claude Team and Enterprise users. Most review times are shortened from about 20 minutes to 6-10 minutes, with speed improved by about 3 times, while discovering more critical bugs with more concise conclusions. Administrators can enable it with one click in Claude Code settings; after installing the GitHub App and selecting repositories, every new PR automatically triggers a review. In addition, eligible Enterprise organizations can claim at least $10,000 in promotional credits by January 15, valid for 30 days after claiming.
Open-source project OpenUI 1.0 has been officially released, allowing large models to generate interactive interfaces such as cards and forms in real time, with a self-developed DSL saving about 67% of tokens and GitHub stars exceeding 10,000.
The open-source project OpenUI (created by thesysdev, under the MIT license) has officially released 1.0, positioned to let LLMs generate interactive interfaces such as cards, charts, forms, and maps in real time to replace pure text answers. The project is model-agnostic and can be paired with local open-source models like Qwen, Gemma, and Llama on Ollama, and the self-developed OpenUI DSL saves about 67% of tokens compared to JSON solutions. Currently, GitHub stars have exceeded 10,000 and downloads have exceeded 6 million.
OpenAI launched the Rosalind Academic program, providing eligible US university research labs with sponsored access to the GPT-Rosalind model for up to 12 months.
OpenAI launched the Rosalind Academic program on October 16, providing eligible US university research labs with sponsored access to the specialized life sciences AI model GPT-Rosalind for a duration of up to 12 months. The model assists researchers in reasoning across scientific literature and biological data, research data is not used for training by default, and labs can submit applications online.
OpenAI announced that ChatGPT Sites Enterprise is generally available (GA), providing advanced capabilities such as US and EU data residency, workspace custom domains, and intranet private tunnels, while removing the limit on the number of sites for Plus/Pro users.
OpenAI announced that ChatGPT Sites is generally available (GA) for Enterprise, and newly created Sites support US and European Union data residency. Enterprise users can build team applications with databases and authentication, and also enjoy advanced capabilities such as workspace custom domains (published under the organization's verified domain) and connecting to intranet APIs through private tunnels. At the same time, Plus/Pro users have the limit on the number of site creations removed, and OpenAI has also launched official tutorials on learn.chatgpt.com covering the entire process from creation, publication, to sharing.
企业级对话代理公司 Bland 推出 iMessage 商用 Agent，企业只需发送链接即可在 30 秒内自动生成智能体，支持聊天内直接收款。
Original · translation pending10 月 16 日，企业级对话代理公司 Bland 宣布支持 iMessage 渠道：企业只需发送一个链接，系统约 30 秒内自动学习业务、接入 API，生成品牌化 iMessage 智能体。该 Agent 以认证商家资料和蓝色气泡发送消息，支持 Tapback、内嵌组件等原生交互，可在聊天内直接收款，并与既有语音、SMS 代理共享同一套 pathway、persona 与长期记忆。
Claude Code CLI has released version 2.1.301, containing a total of 51 changes. The system prompt has tightened security boundaries, explicitly restricting assistance strictly to authorized security tests and refusing unauthorized probing.
Claude Code CLI released version 2.1.301 (just one day after 2.1.300), containing a total of 51 CLI changes and 1 system prompt change. The core change is the tightening of security boundaries: a new paragraph was added to the system prompt, clarifying that Claude will only assist with authorized security tests, and must refuse unauthorized probing and destructive techniques. Other updates include a retry watchdog, vim % jump support, OTel classifier events, etc.
ChatGPT officially announced the introduction of computer use functionality for Chrome users. In chat, it can view and summarize open tabs, and in work mode, it can control tabs to execute operations such as form filling and downloading.
ChatGPT officially announced an upgrade for Chrome users, bringing computer use capabilities into Chrome. In chat, ChatGPT can view user-opened tabs and summarize information; in work mode, it can control Chrome tabs to execute operations such as sending messages, filling out forms, and downloading files, and can also help organize and group open tabs.
PrismML has released the open-source ternary model series Bonsai 2 based on Qwen3.5. Among them, the 2B size is only 0.43GB, and it can run locally and smoothly in the browser using WebGPU.
PrismML released the open-source ternary model series Bonsai 2 based on Qwen3.5, offering three sizes: 2B, 4B, and 9B, with weights of only 0.43GB, 0.95GB, and 2.0GB, which have been listed on Hugging Face in GGUF and MLX 2bit versions. The 2B version can run locally in the browser with the help of custom WebGPU kernels, measured up to about 500 tok/s on the M4 Max, focusing on local assistant scenarios for daily devices such as mobile phones and PCs.
Anthropic announced that Opus 5.5 Fast Mode on the Claude platform is ending its research preview and has been removed from the waitlist, with speed increases of up to 2.5 times and pricing announced.
Anthropic announced that Opus 5.5 fast mode on the Claude platform is no longer in research preview and has been removed from the waitlist. The mode is up to 2.5 times faster, priced at $8 / $40 per million tokens, and is now available on Microsoft Foundry (Azure-hosted), with Google Cloud support coming soon. Fast mode in Claude Code remains unchanged and consumes additional usage credits. Documentation can be found at https://platform.claude.com/docs/en/build-with-claude/fast-mode
Standard Bots 发布 AI 平台，允许用户仅通过 iPad 和自然语言为机械臂编程，降低了机器人操作的门槛。
Original · translation pendingStandard Bots 发布 Standard Bots AI 平台，CEO Evan Beard 在访谈中称之为「给机器人做 vibe coding」——用户只需一台 iPad，就能用自然语言为机械臂编程，宣称可为公司节省数百万美元成本。该发布显著降低了机器人编程的门槛，让非专业开发者也能操控机械臂。
OpenAI 推出 GPT-Live-1.1 语音模型更新并全量推送。新版本新增六种自然声音，移除了会话中烦人的等待环节，非英语语言表现明显提升。
Original · translation pendingOpenAI 于 10 月 16 日官方推出 GPT-Live-1.1 语音模型更新，覆盖 ChatGPT、Codex 和 dots，当日面向所有用户全量推送。新版本新增法语、德语、印地语、墨西哥西班牙语、韩语和日语声音，口音更自然，并移除了语音会话中令人困扰的「checking...」等待环节。官方称新模型全面优于 gpt-live-1，非英语语言表现明显提升，并已放出完整发布说明供开发者查阅。
fal released the Worlds platform, aggregating cutting-edge world models and integrating LingBot to allow users to build characters, explore environments, direct scenes, and receive real-time responses in real time.
fal released the Worlds platform, providing a new way to create and interact with AI-generated worlds in real time. Users can build characters, explore environments, direct scenes, and observe the world's real-time response to their actions. The platform aggregates the latest cutting-edge world models and has integrated related capabilities such as LingBot, combining world generation with agent interaction.
Vercel announced comprehensive backend capability upgrades at the Ship conference, launching Python GA, Containers, and Daemons support for running asynchronous tasks.
Vercel announced the official rollout of its backend capabilities at the Ship conference. After ten months of development, Python support is now generally available (GA), natively compatible with frameworks such as FastAPI and Flask, and built on standard CPython and PyPI; additionally, Containers were introduced to support more runtimes and system libraries, along with Daemons for running asynchronous tasks.
Krea 开放第二款自研图像编辑模型 Krea 3 的 Beta 测试，宣称端到端 2K 编辑 8 秒内完成，并公布了相应的 API 定价。
Original · translation pendingKrea 正式开放自研图像编辑模型 Krea 3 的 Beta 测试，这是继 5 月 Krea 2 之后第二款完全自研模型。官方称端到端 2K 分辨率编辑 8 秒内完成，质量对标 Nano Banana Pro 与 GPT Image 2.5，定价 1K 每张 $0.03、2K 每张 $0.06。
Datadog 发布 Scout 1.0 模型，用于驱动其 AI SRE 智能体 Bits Investigation 处理复杂故障排查。
Original · translation pendingDatadog 发布 Scout 1.0，这是其首个针对生产事故调查做后训练的 SRE 专用模型，基于 GLM-5.3 Flash 微调，用于驱动 AI SRE 智能体 Bits Investigation，处理复杂故障排查和夜间告警。在公开基准 SREGym 测试中表现与最强前沿模型相当，而成本仅为其二分之一。
Microsoft AI 宣布完成 MAI-Base-2 预训练，这是一个 2.8T 参数的 MoE 基础模型。
Original · translation pendingMicrosoft AI 宣布完成 MAI-Base-2 预训练，这是一个 800 亿激活、2.8T 总参数的 MoE 基础模型，在 held-out 评测中 BPB 表现优异，预训练 FLOPs 效率显著提升。
Perplexity 以 Apache 2.0 协议开源 SPACE 智能体代码执行沙箱平台，驱动其全部任务执行并用于安全与强化学习训练。
Original · translation pendingPerplexity 以 Apache 2.0 协议开源 SPACE（Sandboxed Platform for Agentic Code Execution），这是其 Agentic Runtime Platform 首个组件，也是驱动 Perplexity Computer 全部任务执行的生产级基础设施，内部同时用于安全 agentic 运行时与 RL 训练。CEO Aravind Srinivas 亲自发布，并同步公开安全研究《Escaping SPACE: Part I》。
Liquid AI 发布决策模型 d1-ultra，不生成文本 token 直接输出校准概率，在决策指数登顶并按输入计费。
Original · translation pending10 月 16 日，Liquid AI 发布多模态「决策模型」d1-ultra：接收文本与图片输入后，单次前向直接为是非题、多标签选择和量表问题输出校准概率，不生成输出 token，仅按输入计费 $0.042/百万 token。该模型在 Artificial Analysis Decision Index 以 65 分登顶，领先第二名至少 8 分。
Meta open-sources the Muse Spark Open flagship large model, featuring 816B parameters and a million-token context window.
Meta officially open-sourced Muse Spark Open, featuring a total of 816B parameters, 42B active parameters, a context window reaching 1 million tokens, and native support for text, image, video, and audio multimodality. The model focuses on coding and agentic tasks, can run on two Mac Studio machines, has been listed on seven API platforms, and supports inference engines such as vLLM/SGLang.
ARC Prize verifies Google Gemini 4 Argon achieves a score of 97.5% on the ARC-AGI-3 benchmark.
The official ARC Prize verified that Google Gemini 4 Argon achieved a score of 97.5% on the ARC-AGI-3 benchmark, becoming the second laboratory to break 95% on this benchmark, with a verification score approaching GPT-6 Astra at a lower cost.
Amazon launches Bee Frontier and Bee Eclipse wearable devices, supporting continuous conversation listening, automatic summary generation, and integration with Alexa+.
Amazon launched the Bee series of personal AI wearable devices: the wrist-worn/clip-on Bee Frontier and the phone-back-attaching Bee Eclipse. The devices continuously listen to conversations, automatically generate summaries, to-dos, and reminders, and integrate with Alexa+. The Bee brand, acquired by Amazon, returns through this, sparking discussions on privacy controversies and ecosystem lock-in issues.
Cloudflare open-sources Void v1.1.11, providing a full-stack SDK, CLI, and self-hostable platform supporting one-click build and deployment.
VoidZero (which has joined Cloudflare) released and fully open-sourced Void v1.1.11 on October 15, providing a zero-configuration full-stack SDK, CLI, and self-hostable deployment platform three-piece suite in the form of a single npm package + Vite plugin. A single command `void deploy` completes the build, D1 database migration, and launch on Cloudflare Workers. Evan You stated that the original commercialization plan was abandoned due to the acquisition.
The Verge combined with security researchers tested and found that Energize Lab's Eiliko robot toy stores a large amount of user conversations and children's photos on unprotected servers, involving approximately 68000 devices.
The Verge reporter and hacker Sammy Azdoufal tested and found that Energize Lab's Eiliko robot toy stores user conversations, user profiles, children's photos, and audio on unprotected servers, accessible to approximately 68000 devices and 13 million conversations, of which over 31000 children's profiles are from the United States.
StepFun's 600B sparse MoE model Step 5 Preview has launched on FeatherlessAI, serving in FP8 precision and focusing on long-horizon agent work.
StepFun forwarded news from FeatherlessAI that Step 5 Preview has launched on the platform, serving in FP8 precision. This open-weight model features a 600B sparse MoE architecture with 27B active parameters per token, geared toward long-horizon agent work, with frontier-level performance in software engineering and professional knowledge work (especially finance).
Tessl has released Verifiers, allowing rules to be stored in repositories as JSON files, checked line by line by an LLM judge on each code change, and used locally, in pre-commit, or in CI.
Tessl has released Verifiers, where rules are stored in repositories as JSON files. On each code change, an LLM judge provides a pass or fail for each item, and it can be used like a linter or test locally, in pre-commit, or in CI.
快手旗下 Kling 4.0 视频生成模型开启早期访问。实测表明其支持 Omni Reference 单次注入多张参考图、多镜头调度及动态硬切表现。
Original · translation pending快手旗下 Kling 4.0 视频生成模型发布并开放早期访问。博主 techhalla 等实测显示：Omni Reference 可单次注入 6 张参考图并保持角色一致；单条 prompt 可生成三机位镜头；Dynamic Motions 支持 7 秒 5 个硬切动作戏不崩；另有创作者验证原生 30 秒 1080p 单次生成能力。
NVIDIA has open-sourced the Nemotron 3.5 Super VL multimodal model and the Bolt series of small models, forming a family release spanning from cloud to edge.
NVIDIA released and open-sourced the multimodal model Nemotron 3.5 Super VL, combining visual and text understanding into a single MoE model (120B total parameters, 12B active), with weights, datasets, and training recipes fully disclosed, focusing on long-running agent tasks; on the same day, it introduced the Bolt series of small models for local and mobile use, forming a family release spanning from cloud to edge.
Black Forest Labs 开源 5.6B 参数的 FLUX.3 Image Klein 图像模型，可在 8GB 显存的消费级 GPU 上运行，支持原生 4K 输出与多图组合编辑。
Original · translation pendingBlack Forest Labs 于 10 月 15 日开源 5.6B 参数图像模型 FLUX.3 Image Klein，宣称达到大前沿模型质量，可在单张消费级 GPU（约 8GB 显存）上运行，支持原生 4K 输出、边界框控制与像素级精确编辑，单一模型统一文生图与最多 10 张参考图的编辑组合。官方称内部评测文生图与多参考生成超越 Qwen Image 2.1，Diffusers 与 ComfyUI 均提供 Day-0 支持，权重已上架 Hugging Face 并登顶趋势榜。
用户对 Krea 2 Turbo 进行 77 个族裔的系统测试，发现未加衣物描述时默认生成高比例裸体，详细提示词可有效控制服装与背景。
Original · translation pendingReddit 用户 matcheal 用同一 seed 对 Krea 2 Turbo 做了 77 个族裔 × 7 种 prompt 的系统测试，发现模型默认先验极强。全身 prompt 不写任何衣物词时，77 名女性中 34 人生成裸体或半裸；详细 prompt 可准确控制服装、背景与表情，不写的部分则会被模型默认值填充。
开发者公开 LongTakeNoCuts 3分钟一镜到底视频生成案例，通过遮罩 refine 与 seed 控制解决移动镜头中人物崩脸及衔接问题。
Original · translation pending作者用自制的 LongTakeNoCuts 工作流生成 3 分钟一镜到底式视频案例：1 张角色参考图 + ref2va 工作流，hybrid model b20-49，kitchen attention、无 LoRA、10 步，SUNO 生成歌曲，19 段 10 秒片段，CapCut 拼接调色。关键技巧包括用角色遮罩只对人物做 refine 避免镜头移动时崩脸，以及通过重掷并锁定 seed 解决 latent 融合不佳等问题。
