Google DeepMind releases research preview of world model Genie 4 with real-time scene editing
要闻
Google DeepMind has released a research preview of Genie 4 to trusted testers, supporting real-time scene changes through text and synchronized audio generation. The company says continuous interaction is limited to a few minutes, with details and objects more likely to drift as a simulation runs longer.
Google DeepMind is currently offering its world model Genie 4 to trusted testers as a research preview, with access to follow for Google AI Ultra subscribers in Project Genie and early-access partners using Gemini API and Google AI Studio. According to the company, Genie 4 uses Gemini Omni to depict appearances and behavior resembling the real world, while other Gemini models maintain simulation coherence; users can adjust the weather or camera, or add characters and objects, through text prompts at any time. The company says the model can simulate physical effects such as aerodynamics and collisions and produce natural character movements. It also describes Genie 4 as its first real-time world model with native audio generation, producing ambient sound, music, and character dialogue in sync. Gemini 3.8 Flash breaks goals within the simulation into steps and checks the visuals to verify progress. The company also says continuous interaction is limited to a few minutes, and details and objects become more likely to drift the longer a simulation runs.
Cursor offers Annotate in version 3.24.18 and later, letting users mark changes directly on a running app for an Agent to implement in code. Notes include details of the selected target and a screenshot, with batch sending also supported.
Cursor has made Annotate available in Design Mode in version 3.24.18 and later; the source does not specify a launch date. It is accessed through the browser toolbar inside the Agents Window, while earlier versions retain a pen icon labeled Design Mode. According to the official documentation, users can select one or multiple elements, mark an area, edit page text directly, or change the selected element’s CSS properties. Sent notes include details of the selected target and a screenshot to help the Agent locate and modify the code. Batch edits is off by default and, when enabled, allows multiple notes to be sent together, but text edits always wait for Send regardless of that setting. Users can submit follow-up notes while the Agent is working; if the development server supports hot reload, completed changes appear in the running app. Canvas tabs retain the older entry point, send notes directly to chat, and support neither Batch edits nor in-place text editing.
Google begins rolling out Gemini 4 Argon to Antigravity IDE users
开发生态
Google’s Gemini desktop page lists one-click access to Antigravity from the app for building apps, writing code and prototyping. Google says the related macOS features are available, with more native features coming to Windows gradually; no dates are specified.
Google describes features already available on macOS on its Gemini desktop page, including one-click access to Antigravity, without specifying a launch date. According to Google, users can open Antigravity from the Gemini app to build apps, write code and prototype ideas. They can also share their screen for assistance based on its contents, or use voice input to enter and revise text across different apps. The Tasks feature runs multistep work in the background, including organizing local files and compiling their contents into Google Docs or Sheets, with users approving each action.
According to the page, the Gemini desktop app is available in all languages and countries supported by the Gemini app, though feature availability varies, and more native features will gradually come to Windows. The Windows version requires Windows 10 or later and supports x64 and ARM64. The macOS version requires a MacBook with Apple silicon running macOS Sequoia (15.0) or later. Chat history and memory sync across devices signed in to the same Google account. Some features are limited to users aged 18 or older with a Google AI subscription; setup is required, and availability may change.
蚂蚁百灵 open-sources Ling-3.1-flash-1019 and its base model
模型发布
Ant Bailing has open-sourced the previously released Ling-3.1-flash-1019 and made a base model available for further training. The version has 560 billion total parameters, supports a 1 million-token context, and has weights available through Hugging Face.
Ant Bailing has open-sourced the previously released Ling-3.1-flash-1019 alongside Ling-3.1-flash-base, a base model intended for further training. Ling-3.1-flash-1019 activates 25 billion parameters per token out of 560 billion total parameters, supports a 1 million-token context, and has model weights available through Hugging Face. According to the official account, the upgraded version scored 44 on the Artificial Analysis Intelligence Index v4.3.2, compared with 41 for its predecessor, Ling-3.1-flash; its AutomationBench-AA score rose from 62% to 68%.
微软 releases MAI-Image-2.7-Preview image editing model
模型发布
Microsoft has released a preview of its MAI-Image-2.7 image editing model. According to the company, it ranks No. 3 on Arena Edit and No. 2 on Artificial Analysis’s image editing leaderboard, outperforming 2.6 across all 10 use cases and all 7 editing actions evaluated by the latter.
Microsoft has released the MAI-Image-2.7 image editing model in preview, with access available through Arena, Artificial Analysis and MAI Playground. Editing is available first, with text-to-image generation to follow; the source provides no specific date. The model supports changes to the style and lighting of an entire scene, modifications to text and symbols within images, and restoration and enhancement of existing assets. It also supports edits requiring an understanding of relationships between objects, lighting and environments.
According to Microsoft, MAI-Image-2.7 outperforms 2.6 across all 10 use cases and all 7 editing actions evaluated by Artificial Analysis, reaching frontier performance in enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing. In scene and style, its score is 24 Elo ahead of the next-best model. The company also reports that the model ranks No. 3 on Arena Edit, placing Microsoft AI as the No. 2 lab, and No. 2 on Artificial Analysis’s image editing leaderboard.
Microsoft releases MAI Cyber 1.1 Flash security model through MDASH
模型发布
Microsoft is offering the MAI Cyber 1.1 Flash security model to verified customers through MDASH. The company says internal evaluations show performance comparable to MAI Cyber 1 Flash, with an 18% reduction in cost per scan.
Microsoft has made MAI Cyber 1.1 Flash available to verified customers through MDASH, with safeguards including clear usage policies and ongoing monitoring. According to the company, the new model handles more of the vulnerability-discovery work itself, aiming to reduce operating costs across the system while preserving performance. On the same internal MDASH vulnerability-scanning evaluation suite, it delivered performance comparable to MAI Cyber 1 Flash at an 18% lower cost per scan. The company also reported that after customers run their first scan with the MAI Cyber 1 Flash profile, nearly 9 in 10 subsequent scans use the MAI profile rather than the OpenAI-only model profile. MAI Cyber currently accounts for almost 50% of all MDASH scans, with more than thousands of scans run by hundreds of customers.
商汤 open-sources SenseNova 6.8 Code2Video video model
模型发布
SenseTime has open-sourced SenseNova 6.8 Code2Video, which creates videos by generating executable code and revises them within the same project based on feedback. According to the official description, it supports a 512K context and can sustain runs exceeding 10 hours and 2,000 turns.
SenseTime has released the open-source video model SenseNova 6.8 Code2Video and made its weights available on Hugging Face; the source does not specify a release date. Built on Qwen3.8-27B, the model constructs and renders videos by writing and executing code, currently focusing on short videos implemented entirely in code. Its production workflow includes writing specifications, rendering individual frames, reviewing images, and revising code. Users can provide feedback in the same session or resume the original project later. According to the official description, the model supports a 512K context, can sustain runs exceeding 10 hours and 2,000 turns, and can revise code in response to feedback without changing parts not designated for adjustment. The repository also provides a code-video skill covering visual review, verification of delivered files, and eight categories of production guides. The runtime environment requires Claude Code, Node.js 18+, Python 3, ffmpeg/ffprobe, and fonts for the video's language. The team plans to scale the model further and extend its capabilities to editing, compositing, animation, and 3D content creation.
百度 open-sources Unlimited OCR 2 with support for long-document parsing
模型发布
Baidu has open-sourced Unlimited OCR 2 on GitHub, supporting the parsing of an entire document’s page images in a single request. The project offers Transformers and SGLang execution paths and says multi-page parsing can use context across pages.
Baidu has open-sourced Unlimited OCR 2 on GitHub with support for long-document parsing. The project provides a native Transformers path for quick tests and recommends SGLang for serving. According to the project, infer_multi processes document images supplied in page order within a single request, allowing the model to use cross-page context; each page is encoded as one global view, and the output preserves one <PAGE> marker per page in input order. The single-page interface, infer, uses dynamic tiling by default. Deployment uses a customized SGLang wheel and PyMuPDF to convert PDFs into images, while the service exposes an OpenAI-compatible API. The batch client infer.py is used for MPDocBench / MarathonDocBench, sending each document as a single request and saving the raw output, which utils/process_unlimitedOCR2.py then converts into evaluation Markdown. The project also provides a replay showing a full-length book parsed in a single request.
OpenBMB open-sources MiniCPM-OCR, including a 1.3B parsing model and a 2B scoring model
模型发布
OpenBMB has open-sourced the MiniCPM-OCR series, comprising a 1.3B model for Chinese and English document parsing and a 2B model for scoring OCR output quality. The code, documentation, and model weights are all released under Apache 2.0.
OpenBMB has released and open-sourced MiniCPM-OCR, a two-model series for document parsing and OCR quality scoring; the source does not specify a release date. The parsing model, MiniCPM-OCR 1.0, has 1.3B parameters and supports Chinese and English documents and PDF parsing. According to OpenBMB, it scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. The scoring model, MiniCPM-OCR Jev, has 2B parameters and can assign a quality score from 0 to 1 based on an image and its OCR output. It also supports evaluating output from other parsers, including GLM-OCR and PaddleOCR-VL-1.6. The series supports vLLM, SGLang, llama.cpp, and Ollama, with its code, documentation, and model weights all available under Apache 2.0.
Liquid AI releases LFM3, a family of six 2B-class open-weight multimodal models
模型发布
Liquid AI has released LFM3, initially offering six open-weight models with around 2B parameters each and a shared backbone for text, vision, audio, omni, and decision tasks. Base and post-trained models are available through Hugging Face, Playground, and the Liquid AI API.
Liquid AI has released the LFM3 family and made the weights of its base and post-trained models available at launch. All six initial models derive from LFM3-1.7B-Base and have around 2B parameters each. Four handle text, vision, audio, and omni inputs respectively, while the other two are decision models specialized from the vision and omni models. They are available through Hugging Face, Playground, and the Liquid AI API. According to the company, the family targets specialized tasks and supports local execution and fine-tuning. The text model, LFM3-1.7B, has a 256K-token context window and was post-trained for Agent frameworks including Hermes Agent, Pi, and OpenClaw.
On training and architecture, LFM3-1.7B-Base was trained on 22T tokens, with the final 3.5T tokens forming a mid-training stage using curated data that also extended the context window to 256K tokens. Its 24 layers comprise 13 gated short convolution layers, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. The company reports that, in tests using llama.cpp at Q4_0 with a 32K-token context, LFM3-1.7B decoded 1.8 to 2.8 times faster than the next-fastest model on all three tested devices and used the least memory. A full 128K-token context required 1.6 GB of KV cache, compared with 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B.
SparkLLM open-sources Spark-X2.6-1.7B, with post-training improvements to coding and agent capabilities
模型发布
SparkLLM has open-sourced Spark-X2.6-1.7B, retaining its predecessor’s architecture while revising post-training. In the team’s evaluations, the model scored 59.8 on SWE Verified and 85.3 on τ²-bench with thinking set to high.
The SparkLLM team has released Spark-X2.6-1.7B under the Apache 2.0 license, with downloads available on Hugging Face and ModelScope and support for running it through Ollama. It is the second model in the Spark-X2.6 series after Spark-X2.6-4B, retains the architecture of Spark-X2.5-1.7B, and offers three configurable thinking levels: none, low, and high. According to the team, the post-training changes target coding, agent, long-context, multilingual, and reasoning capabilities. In its evaluations, SWE Verified scores rose from 28.3 to 59.8 and τ²-bench scores rose from 65.3 to 85.3, with X2.6 using the high setting.
Mirendil releases its first model, Fern, with support for up to 10M tokens of context
模型发布
Mirendil has released Fern, its first model, with support for up to 10M tokens of context and an initial version named Fern 0.1. The company positions it for focused, high-volume Agent tasks and says new versions will ship frequently.
Mirendil released its first model, Fern, on October 26, 2026, at 8:00 am PT, with the initial version designated Fern 0.1. According to the company, the model was pretrained from scratch, supports a context window of up to 10M tokens, and underwent a short round of post-training for its initial release. Mirendil describes Fern as its most efficient model, positions it for focused, high-volume Agent tasks, and says new versions will ship frequently.
ChatGPT has announced two independent plugins, iCloud Mail and iCloud Calendar, for all subscription plans. They support email and calendar operations and require an Apple app-specific password to connect.
ChatGPT has announced the launch of two independent plugins, iCloud Mail and iCloud Calendar, across web, desktop and iOS for all ChatGPT subscription plans, with actual availability determined by workspace app control settings. According to the official announcement, users can use the plugins to ask ChatGPT to find emails, organize their inbox, check their next scheduled event or update calendar events; both plugins also include read-only contact lookup. Connecting requires an Apple app-specific password.
Arena adds Connectors to Agent Mode, enabling connections to apps such as GitHub
产品应用
Arena has made Connectors in Agent Mode available free to signed-in users, with support for eight listed apps, including GitHub and Google Drive, allowing the agent to read service content and perform actions directly in chat.
Arena has launched Connectors in Agent Mode, and signed-in users can now access the feature free at arena.ai/agent; the source does not specify a launch date. Once users connect their existing apps, the agent can read content from those services and perform actions directly in chat. The listed supported apps include GitHub, Google Drive, Notion, Linear, Gmail, Supabase, Vercel, and Dropbox. According to Arena, the feature enables users to hand more complex tasks to the agent for end-to-end execution.
Nous Research revises Hermes Agent subscription plans
产品应用
Nous Research has revised its Hermes Agent subscription plans, including a cloud Agent by default in new plans and replacing monthly usage allocations with daily and weekly limits. Subscribers can switch at any time and connect to the Agent through messaging apps without keeping a laptop running.
Nous Research has announced changes to Hermes Agent subscriptions, replacing a single monthly usage allocation with daily and weekly limits; subscribers can now switch to the new plans at any time. According to the company, each new plan includes a ready-to-use cloud Agent by default. Nous Research’s Teknium said that paid Nous Portal subscriptions now include a Hermes Agent instance and a computer for it to use. Users can connect to the Agent through various messaging apps without keeping their laptops running, and can log in with one click in the desktop app.
千问 launches study features for finding, solving, and practicing problems
产品应用
Qianwen has launched learning features covering exam-paper searches, photo-based question support and homework grading. According to the service, users can download and print exam papers for free, while everyday mistakes are automatically saved in learning records and a weakness analysis identifies repeatedly missed concepts.
Qianwen has launched learning features that bring exam-paper searches, problem explanations and mistake tracking into a conversational learning workflow; the source does not specify a launch date. According to the service, users can find exam papers and download and print them for free by entering their region, grade level, subject and exam-paper keywords in the chat box. For problem solving, users can submit a photo through “拍题答疑,” and the system provides solution steps based on grading criteria. If a step remains unclear, users can move to “小讲堂,” where an AI teacher breaks it down into smaller steps. Everyday mistakes detected by “作业批改” are automatically saved in the upgraded “学习档案.” Its “薄弱点分析” lists repeatedly missed concepts for targeted practice.
即梦 AI launches a short-drama studio integrating creation and remake features
产品应用
Jimeng AI has launched a short-drama studio that brings scripts, art design, storyboards, post-production, and remaking and redrawing into one tool. Limited-time free allowances include one script creation and one script analysis, plus 15 seconds each of intelligent previews and white-model remaking and redrawing.
Jimeng AI has officially launched a short-drama studio integrating functions from script creation to post-production adjustments and remaking and redrawing; the source does not specify a launch date. According to the company, the studio uses its in-house script model DreamScript, image model Anycook, and storyboard model DreamShot to parse long texts, generate characters and scenes, and coordinate storyboard adjustments. Post-production tools include clip reshoots, audio reconstruction, and intelligent shot cutting, while remaking and redrawing regenerates visuals according to the selected style, aspect ratio, and target region. The platform currently offers limited-time free allowances comprising one script creation, one script analysis, 15 seconds of intelligent previews, and 15 seconds of white-model remaking and redrawing.
WeChat Input Method has released version 4.0.0 with an AI voice recorder, available within the App on iOS and Android and also available on Windows. The feature transcribes audio during recording, uses AI to summarize the content, and generates shareable, interactive web pages.
WeChat Input Method has released version 4.0.0, adding an AI voice recorder now available on iOS, Android, and Windows. Users on iOS and Android can access the feature directly within the App, and the Windows version also supports it. According to the official description, the AI voice recorder converts speech to text while recording and uses AI to summarize the recorded content; the generated web pages can be shared and support interaction. Its outputs include not only transcripts of recordings but also content summaries and web pages.
蚂蚁集团 releases 灵光 2.0 with ongoing task follow-up in the background
产品应用
Ant Group has released Lingguang 2.0 and made it fully available, allowing authorized tasks to continue after users leave a conversation. The update covers cross-session memory, long-running task scheduling and tool connectivity, while letting users review, modify or revoke permissions required for Agent execution.
As of October 28, 2026, Ant Group has released Lingguang 2.0 and made the product fully available. The update covers cross-session memory, long-running task scheduling and tool connectivity. According to the company, Lingguang runs on an independent cloud host and identifies needs after receiving user authorization. It can continue executing tasks after users leave a conversation and adjust arrangements based on new information. For service connectivity, the product supports connections to Alipay, DingTalk and Feishu, as well as calls to tools such as MCP and CLI. Lingguang 2.0 also integrates Ant APASS, incorporating identity verification, intent security and privacy protection into the Agent execution process; users can review, modify or revoke authorization.
Claude also takes on ETH after OpenAI publishes a disproof
技术与洞察
Anthropic’s formal-math repository has released a 3-SAT algorithm discovered by an internal research Claude model, along with Lean 4 formalization materials, claiming a subexponential-time refutation of ETH. The result concerns asymptotic complexity only and says nothing about practical solving performance.
Anthropic has published a research result in its formal-math repository that it claims refutes the Exponential Time Hypothesis (ETH); the page does not specify a publication date. According to its description, a deterministic algorithm discovered by an internal research Claude model solves 3-SAT with n variables in 2^{o(n)} time. It also solves the harder problem #3-SAT, which counts satisfying assignments, in 2^{O(n log log n / √log n)} time. Anthropic states that the result concerns asymptotic complexity only, not practical solving performance.
The repository provides the paper and Lean 4 formalization materials written by AI models, with versions pinned to Lean v4.33.1 and Mathlib v4.33.1. According to its description, CI compares theorem statements and rechecks the proofs using nanoda, a second, independently implemented kernel. However, no machine checks whether the formalized statements match what the paper says; readers are left to verify that correspondence. Anthropic also states that Claude Opus 5.5 at high effort can rederive, from a condensed description, the conclusion that 3-SAT can be decided in O(2^{δn}) time for every δ > 0, but not the sharper bounds above. The repository is published as a research artifact, is not maintained, and does not accept contributions.
Anthropic hiring manager says the team no longer uses traditional coding interviews
技术与洞察
An Anthropic hiring manager has stopped using traditional coding interviews in his own hiring, arguing that Claude has undermined their value as an assessment tool. He also outlined five practices for less-experienced engineers, centered on business goals and working prototypes that demonstrate an idea.
An Anthropic hiring manager wrote that he stopped requiring candidates to take traditional coding interviews during the month he wrote the article; the source does not specify the month or year. This describes his own hiring practice, rather than announcing an Anthropic-wide policy. According to his account, small coding exercises once provided a limited indication of how candidates would perform on his teams, but Claude now performs better on small React challenges and assembly questions, while the team's most impactful builders succeed regardless of their scores on such tests. He argues that AI makes engineers' impact less constrained by the years they have spent accumulating knowledge of frameworks and libraries, and defines his intended audience as engineers lacking social capital, status, or power. He outlines five practices: prioritize understanding business goals rather than simply processing the backlog; develop ambitious ideas that could advance those goals; build a working prototype over a weekend; use the prototype to communicate just one central idea; and use a newcomer's perspective to identify problems in user experience or infrastructure. He says he has seen individuals and small teams use prototypes to bypass the usual process of writing PRDs, securing support, and coordinating time with designers or product managers.
Cognition releases FrontierCode 2.0 to test whether models can write code worth merging
技术与洞察
Cognition released FrontierCode 2.0 to assess engineering under incomplete instructions through code mergeability. The top model, Opus 5.5, scored 45%; solutions that fail blocking criteria receive zero.
Cognition released FrontierCode 2.0, a coding benchmark focused on whether models can deliver mergeable code from incomplete requirements. The company says it tests not only explicit requirements but also the ability to infer constraints from a codebase, including reusing existing abstractions, following repository conventions, and writing tests that catch bugs. Scores are weighted aggregates of grading criteria, with solutions that fail blocking criteria receiving zero. In the published results, Opus 5.5 ranked first at 45%. Cost is measured as mean USD spending per rollout, using list API prices with public discounts included.
According to the company, more than 20 open-source developers built evaluations from repositories they maintain, each spending over 40 hours per task. Compared with FrontierCode 1.0, the new version expands task diversity, grading criteria, and task scope, while strengthening reviews of task prompts, reference solutions, and grading criteria. The share of tasks whose reference solutions differ from upstream PRs rose from 60% to 80%. Among blocking criteria, the number of thoroughness criteria, which test whether models infer implicit correctness requirements, reached 3x the count in FrontierCode 1.1. The company also notes that increasing reasoning effort does not necessarily improve scores, because models may write more code and modify more files while exceeding the required scope.
Qwen team open-sources BabelArena to evaluate agents across 23 natural languages
技术与洞察
The Qwen team has open-sourced BabelArena, expanding 702 standard tasks from four benchmark families into 23 natural languages, for a total of 16,146 task instances to evaluate multilingual Agent performance.
The Qwen team has open-sourced BabelArena, a multilingual Agent benchmark covering 13 domains; the source does not specify a release date. Built from 702 standard tasks across four benchmark families—τ³-Bench, VitaBench, DeepPlanning, and ClawEval—it provides versions in 23 natural languages, totaling 16,146 task instances. Localization covers prompts, tasks, tool descriptions, and environment content, with languages including English, Chinese, and Arabic.
The study evaluated 5 frontier models. According to the Qwen team, no model led across all four benchmark families, and Pass^1, Pass^3, and language consistency were all lower in low-resource languages than in high-resource languages. The paper also reports that Agent input token counts in low-resource languages can reach approximately twice those in English.
Axiom says AI completed a Lean formalization of the classification theorem for finite simple groups in 24 days
技术与洞察
Axiom’s project to formalize the classification of finite simple groups has reported a group theory result: the team found counterexamples to a closure assertion used in a published route to the Unbalanced Group Theorem and presented an alternative argument for the simple-group case in a 35-page paper.
Axiom’s website presents a group theory paper arising from its project to formalize the classification of finite simple groups, reporting counterexamples in an existing proof route and an alternative argument for the simple-group case; the supplied material gives no publication date. The paper is identified as Hexagon 2610.00575 and runs to 35 pages. According to the website, while formalizing the first-generation proof, the team found counterexamples to a closure assertion used in a published route to the Unbalanced Group Theorem. The paper explains this obstruction, while its alternative argument builds on Solomon’s 1976 theory of maximal 2-components. The key step is Theorem 2.3, concerning the preservation of simple tops.
