OpenAI officially launches Linux preview of ChatGPT desktop app
要闻
OpenAI announced that a Linux preview of the ChatGPT desktop app is now available, with support for ChatGPT, ChatGPT Work, and Codex features. It can be used on select distributions, including Debian and Fedora, across both x64 and ARM64 architectures.
OpenAI has officially announced that the Linux version of the ChatGPT desktop app has entered preview. The app integrates ChatGPT, ChatGPT Work, and Codex, allowing users to work on projects and use browser workflows on supported Linux systems. The preview currently supports desktop versions of Ubuntu 24.04 LTS and 26.04 LTS, Debian 13, and Fedora 43 and 44. Users can install it via .deb or .rpm packages, with both x64 and ARM64 architectures supported.
SpaceXAI launches Grok Bot as Elon Musk teases Grok 4.6 for this week
要闻
SpaceXAI has released an early test version of Grok Bot, an Agent with its own cloud computer that can log in to existing tools and platforms without public APIs and complete multistep tasks end to end. Desktop and iOS versions are currently available to subscribers of plans including SuperGrok Heavy. Elon Musk said testing will expand after basic issues are fixed and that Grok 4.6 is planned for release this week.
SpaceXAI has officially launched an early beta of Grok Bot, an AI teammate capable of performing real-world work. Grok Bot has its own cloud computer and can log in to users’ existing tools and apps to complete multi-step tasks end to end. According to the company, users can assign tasks to a Bot just as they would message a colleague. The Bot can remember conversations, learn user preferences, and improve its performance through repeated collaboration. Users can run multiple Bots simultaneously; they can communicate with one another, share context, and coordinate work autonomously in group chats. Grok Bot is currently available to SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers on desktop and iOS, while enterprise users can join the waitlist. Elon Musk said the beta will be expanded after its basic early-stage issues are resolved, and Grok 4.6 is scheduled for release later this week.
Researchers say major large models are vulnerable to cross-model jailbreaks that extract hidden reasoning in plaintext
要闻
Researchers recently published a paper claiming that vulnerabilities in proprietary large language model APIs can be exploited by injecting encrypted reasoning blocks from a stronger model into a weaker one to jailbreak it and extract hidden reasoning in plaintext. The researchers said they responsibly disclosed the issue to the companies concerned.
A research team has published a paper reporting a vulnerability in major large-model APIs that allows encrypted reasoning blocks to be reused across different sessions, users, and models from the same provider. By injecting encrypted reasoning blocks from powerful models into weaker compatible models and jailbreaking them, the researchers successfully forced the weaker models to decrypt and output the powerful models’ original plaintext reasoning verbatim. The method worked on models from Anthropic, OpenAI, and Google. The researchers said they had notified the AI labs through responsible disclosure procedures. The labs have patched multiple issues caused by the vulnerability and are continuing to implement fixes. The study also identified sensitive data, including API keys and passwords, by scanning publicly available reasoning traces, and observed clear discrepancies between models’ actual internal reasoning processes and the summaries shown to users.
Bilibili IndexTeam releases open-source IndexTTS-2.5 voice-cloning model
要闻
Bilibili has officially released the weights for IndexTTS-2.5, a zero-shot voice-cloning model supporting cross-lingual transfer among Chinese, English, Japanese, Spanish, and Arabic. Its redesigned architecture improves inference efficiency and enhances pronunciation and emotion control, while also supporting speech-speed adjustment from 0.5x to 2x. The code and weights are now available on the relevant platforms.
Bilibili’s Index SpeechTeam has officially open-sourced IndexTTS-2.5, a zero-shot text-to-speech model. The model has 0.8B parameters and supports voice cloning and cross-lingual transfer across Chinese, English, Japanese, Spanish, and Arabic. Compared with the previous generation, IndexTTS-2.5 has reworked its inference architecture to improve runtime efficiency and added speech-rate controls ranging from 0.5× to 2.0×. The project’s complete code, model weights, and paper are now available on the relevant platforms under the Bilibili Model License.
智谱's ZCode reaches 1 million users and resets quotas for all subscribers
要闻
智谱 announced that its Harness product ZCode has surpassed one million users and that quotas for all GLM Coding Plan users were reset at 1 p.m. on August 11. The company also introduced features including idle-time tasks that run automatically during off-peak hours without consuming quota.
Zhipu has announced that ZCode, its Harness product, has surpassed one million users and that the quotas of all GLM Coding Plan users were reset at 13:00 on August 11. The company also announced a major ZCode upgrade, introducing Idle Tasks that run automatically during off-peak hours without consuming quota, Goal autonomous delivery mode, parallel collaboration through Subagents, Remote Control for mobile control, and other features.
ChatGPT desktop app and Codex CLI add imports from other Agents with automatic syncing
要闻
OpenAI has introduced an Agent import feature that allows the ChatGPT desktop app and Codex CLI to import configurations, projects, and recent work from Claude Code, Claude Cowork, and Cursor. The ChatGPT desktop app can also enable automatic updates and syncing with the original Agent.
OpenAI announced that the ChatGPT desktop app and Codex CLI now support importing work from other Agents. The ChatGPT desktop app can import from Claude Code, Claude Cowork, and Cursor, while Codex CLI can import from Claude Code and Cursor. Transferable content includes instruction files, settings, skills, plugins, project folders, MCP server configurations, hooks, subagents, and chat history from the past 30 days. The ChatGPT desktop app also supports automatic updates to keep imported content synchronized with the source Agent.
Google: Gemini surpasses 1 billion monthly active users as Gemma tops 1 billion downloads
要闻
Google CEO Sundar Pichai announced that the Gemini app has surpassed one billion monthly active users, making it the fastest-growing product in Google's history. Downloads of the Gemma model family have also exceeded one billion.
Google CEO Sundar Pichai announced that the Gemini app has surpassed 1 billion monthly active users, making it the fastest-growing product in Google’s history and the company’s 14th product to reach the 1 billion-user milestone. According to Google, 63% of users interact with Gemini by voice, users generate more than 150 million images per day, and over 100 million people use Gemini on iOS. During the same period, Google’s official Google Gemma account confirmed that cumulative downloads of the Gemma model family have also surpassed 1 billion.
LTX has released LTX-2.5, an open-source world model that generates high-fidelity video and audio from text, images, and video. It supports native multi-shot scenes with consistent characters and uses diffusion-based fidelity rendering to improve detail. Its weights are open, and it can run locally with as little as 16GB of VRAM.
LTX.io officially released LTX-2.5, an open-source world model that generates high-fidelity audiovisual content from text, image, and video inputs. The new version introduces native multi-shot generation, diffusion-based fidelity rendering, and a custom Gemma 4 12B text encoder, improving visual detail and prompt adherence. The model is available with open weights, can run locally with as little as 16GB of VRAM, and allows users to fine-tune it on their own data. According to the company, most LoRAs previously trained on LTX-2.3 can run directly on LTX-2.5.
NVIDIA has released Nemotron 3.5 Lightning, an open-source model with a 30B MoE architecture that activates only 3B parameters. Designed for high-frequency execution tasks performed by long-running Agents, it is now available for free on OpenRouter and OpenCode.
NVIDIA released Nemotron 3.5 Lightning, an open-source large language model featuring a hybrid Mamba-2 + MoE + Attention architecture. It has 30B total parameters but activates only 3B parameters per token. Distilled from Nemotron 3 Ultra, the model is designed for the high-frequency execution layer of long-running AI Agents and supports context lengths of up to 1M tokens. Released under the open OpenMDW-1.1 license, its weights, training data, and training recipe are all publicly available. It supports deployment on a range of hardware, including DGX Spark, H100, and GB200, and is also available for free on OpenRouter and OpenCode.
商汤 has launched the multimodal model SenseNova 6.8 Flash Lite Preview. Users can access it early through the SenseNova Token Plan, with a limited-time free allowance of 1,500 uses every five hours. The company expects to release the production version of 6.8 Flash Lite with further performance improvements, along with the 6.8 Flash model, later this month.
The preview version of SenseTime’s new SenseNova multimodal agent model, SenseNova 6.8 Flash Lite Preview, is now officially available for users to try through the SenseNova Token Plan. Designed to be lightweight and agile, the model offers autonomous planning, multi-Agent collaboration, and dynamic course correction, advancing toward “delegative intelligence,” in which users only need to specify a goal and the AI independently completes the entire workflow. Users currently receive a limited-time free quota of 1500 uses every 5 hours. SenseTime said the production version of 6.8 Flash Lite, which will deliver further performance improvements, and the 6.8 Flash model are expected to be released in August.
Microsoft's MAI-Code-1.1-Flash model rolls out to GitHub Copilot
模型发布
Microsoft's MAI-Code-1.1-Flash coding model is rolling out to GitHub Copilot. According to the company, the model adds native visual support, improves CLI and .NET tasks, requires fewer tokens, and delivers faster transmission.
Microsoft’s latest small code model, MAI-Code-1.1-Flash, is now rolling out in GitHub Copilot. Building on its predecessor, the model adds native vision support for image understanding, optimizes CLI and .NET tasks, and delivers improvements in code quality, instruction following, and tool use. According to official data, the model requires 25% fewer tokens to complete tasks and offers a 25% increase in streaming speed.
The Unsloth team has released Unsloth Desktop, an open-source desktop application. It enables users to run and train LLMs and other AI models locally on Mac, Windows, and Linux, with support for chips from NVIDIA, AMD, Intel, and Apple Silicon.
The Unsloth team has officially released the Beta version of its open-source desktop application, Unsloth Desktop, with support for macOS, Windows, and Linux. The application allows users to run and train a wide range of models on local hardware, including LLMs, image and video diffusion models, audio models, MLX, and GGUF, across NVIDIA, AMD, Intel, and Apple Silicon chips, as well as CPU and multi-GPU setups. It includes private web search, deep research, RAG, and MCP, and allows users to connect local models to Agents such as Claude Code and Codex using the unsloth start command.
NVIDIA has released NeMo Switchyard, an open-source model-routing library supporting algorithms such as LLM classifiers, stage routing, and escalation routing. The project is currently in pre-alpha.
NVIDIA has officially released and open-sourced NeMo Switchyard, a model-routing library designed to distribute requests across models in AI Agent workflows. It supports algorithms including LLM classifiers, phase routing, escalation routing, and prefill routers. NVIDIA says the project is available on GitHub, but it is labeled pre-alpha and is not intended for production use.
Anthropic expands Compliance API for Enterprise customers
开发生态
Anthropic has extended the Compliance API to Claude Cowork and Claude Code and opened Beta testing to Enterprise customers. Compliance teams can retrieve conversation content and metadata from both products for audits and forensic investigations without separate integrations.
Anthropic has officially announced that the Claude Compliance API has expanded its coverage to Claude Cowork and Claude Code. Security and compliance teams can now use the existing Compliance API interface to retrieve conversation content and metadata from Cowork’s desktop application, web interface, and mobile app, as well as from the CLI and desktop versions of Claude Code. This new capability is currently in Beta and is available to Claude Enterprise customers.
WorkBuddy adds multi-device syncing and remote control of computer Agents from phones
产品应用
WorkBuddy has introduced multi-device syncing across PCs, its APP, and mini programs. By signing in to the same account, users can remotely control a computer-based Agent from their phone, including while the screen is locked, and collaborate across both devices.
Tencent WorkBuddy has introduced multi-device synchronization, enabling real-time syncing of tasks, conversation history, and outputs across PC, APP, and Mini Program clients. Users can remotely authorize or stop tasks being performed by a desktop Agent from their phones, and can connect a single phone to multiple computers and switch between them. It also supports dual-device collaboration, allowing the mobile device to collect information that the desktop Agent can directly process and extract.
OpenAI launches Agent feature for ChatGPT workspaces
产品应用
OpenAI has launched a workspace Agent feature that lets users create shareable Agents in ChatGPT to handle complex, multistep workflows across tools. The feature is currently in research preview and is available only on ChatGPT Business, Enterprise, Edu, and Teachers subscription plans.
OpenAI has introduced workspace Agents in ChatGPT. Users only need to describe their work, and ChatGPT can help turn it into a functional Agent that can be shared across teams to automate long-running, multi-step workflows such as reviewing prospects, generating reports, and researching vendors. These Agents can pull contextual information from tools including Slack, Google Drive, and Microsoft SharePoint, and, once approved, take actions such as updating tickets or sending messages. The feature is currently in research preview and is available to users on ChatGPT Business, Enterprise, Edu, and Teachers plans.
Manus officially announces return to independent operations, with some user data to be deleted for regulatory reasons
行业动态
Manus announced that it will soon resume operating as an independent company. As part of its separation from Meta, data generated by some users after a specified date will be deleted between August 23 and 24, Singapore time. Affected users must back up their data in advance and will be able to restore it starting August 25.
Manus has issued an open letter to users, announcing that it will soon resume operations as an independent company. According to the official statement, Meta acquired Manus on December 29, 2025. To comply with regulatory requirements in certain jurisdictions, data generated by some users on or after the date of the acquisition will be deleted. Affected users may use the official backup tool to back up their data from now until 07:59 on August 23, 2026 (Singapore Time). The data will be deleted between August 23 and 24 (Singapore Time), and users may restore their data and resume normal use starting at 08:00 on August 25 (Singapore Time). Unaffected users do not need to take any action and may continue using Manus as usual.
Mistral AI launches regional inference endpoints and adds third-party open models
行业动态
Mistral AI has launched regional inference endpoints, allowing customers to choose between running workloads in Europe or the United States to meet data residency requirements. The Mistral platform will also support third-party open models, starting with 智谱's GLM-5.2.
Mistral AI has announced that it is strengthening customers’ AI sovereignty through regional controls, open model choice, and long-term compute commitments. At the inference layer, Mistral Regional Endpoints are now generally available, allowing customers to choose whether to run inference in Europe or the United States to meet data residency, regulatory, and latency requirements. In addition, the Mistral platform will support third-party open models, starting with Zhipu AI’s GLM-5.2. These models will run on the same infrastructure and under the same regional controls as Mistral’s own models.
Former 通义千问 head 林俊旸 founds 语用科技 to enter embodied AI
行业动态
Former 通义千问 head 林俊旸 announced the founding of 语用科技, a company focused on developing next-generation Agents that operate across the digital and physical worlds. 林俊旸 said the company has received investment from firms including 红杉中国 and 高榕创投.
Lin Junyang, the former head of Tongyi Qianwen, has announced the founding of Pragmatik Labs (语用科技) in Shanghai, with a focus on developing next-generation Agents that span the digital and physical worlds. According to the company’s website, it will build digital Agents capable of performing knowledge work, as well as physical or embodied Agents capable of carrying out long-horizon tasks in real-world environments. Lin Junyang confirmed that the company’s latest funding round was co-led by Sequoia China and Gaorong Ventures, with backing from Tencent and the Shanghai Future Industry Fund.
深度求索 registers “Deepseek Harness 团队” WeChat official account
行业动态
Online users recently discovered that DeepSeek had registered a WeChat official account named “Deepseek Harness 团队.” Available information shows that registration was completed in early July, but the account has not published any posts so far.
According to community reports, Beijing DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd. registered a WeChat official account named “Deepseek Harness Team” in early July. The account is currently registered but has not yet published any content.
Google AMIE meets clinical benchmark in simulated consultation test
技术与洞察
Google announced that its medical AI system AMIE can now conduct real-time video consultations. In a randomized study using simulated consultations, AMIE met the benchmark for clinical performance, though it remains a research system.
Google is advancing its medical AI system AMIE to support real-time clinical video consultations. Built on Gemini and Project Astra, the system uses a multi-agent architecture that can interpret visual and auditory cues, guide virtual physical examinations, and perform diagnostic reasoning in real time. In a randomized study involving simulated consultations with patient actors and primary care physicians, clinical evaluators rated AMIE favorably on core clinical competencies, including the comprehensiveness of history-taking, diagnostic accuracy, appropriateness of management, and communication quality. Patient actors also preferred the video experience over text-based chat. AMIE remains a research system, and further study is needed before it can be responsibly deployed in real-world clinical settings.
腾讯混元 has released an Agentic framework called WorldClaw. It can transform text prompts into large-scale, freely explorable 3D open-world scenes composed entirely of editable 3D assets.
Tencent Hunyuan has launched WorldClaw, an Agentic 3D open-world generation framework. Rather than generating videos or Gaussian splats, the framework converts open-ended text prompts into large-scale scenes built from editable, game-ready 3D assets. Through coarse-to-fine Agentic planning, WorldClaw generates rich local details while maintaining globally coherent terrain. The generated scenes support free-viewpoint exploration, object-level editing, and reuse, and can be integrated directly into rendering, animation production, and game-engine workflows.
OpenAI reportedly plans to offer U.S. college students one year of ChatGPT Plus for free
前瞻与传闻
OpenAI is reportedly planning to offer U.S. college students one year of ChatGPT Plus for free. The benefit would cover currently enrolled students at more than 220 participating U.S. institutions and require identity verification through SheerID for activation. The information has not been officially confirmed.
According to a tipster, OpenAI is preparing a new “Back to School” campaign positioning ChatGPT as a tool for exam preparation and learning. As part of the campaign, currently enrolled college students in the United States will be able to claim one year of ChatGPT Plus for free. The offer will be limited to students enrolled at more than 220 participating U.S. colleges and universities. Eligible students will need to verify their student status through verification partner SheerID. This information has not yet been officially confirmed.
字节跳动 forms new top-level “AI数据与安全” unit to tackle data for large models
前瞻与传闻
According to reports, 字节跳动 recently established a new top-level unit called “AI数据与安全.” The unit consolidates several previously separate teams and will primarily provide cross-modal data services for all of 字节跳动's large models.
According to Intelligent Emergence, ByteDance recently established a new first-level division called “AI Data and Security,” operating alongside divisions such as Seed and Flow. It is headed by Wang Yinglei, formerly responsible for platform accountability at TikTok. The division consolidates teams including Global Data, the group-level data platform DMC, and Flow’s AI data platform AIDP. Its core function is to provide cross-modal data services for all of ByteDance’s large models, covering processes such as standards development, sourcing and procurement, synthetic data generation and cleaning, and quality evaluation.
Riot Platforms and Anthropic sign $9.1 billion data center lease agreement
前瞻与传闻
Riot Platforms and Anthropic have reportedly signed a 20-year data center lease agreement worth approximately $9.1 billion. The deal covers 191 megawatts of computing capacity and includes two five-year renewal options, bringing its total potential value to about $16.1 billion.
Riot Platforms previously announced in its quarterly earnings report that it had signed a 20-year data center lease agreement with a “frontier AI lab” to provide 191 megawatts of critical IT capacity at its Rockdale campus in Texas. The initial contract term is expected to generate approximately US$9.1 billion in revenue. Bloomberg, citing people familiar with the matter, reported that the tenant is Anthropic, though neither Riot nor Anthropic has formally commented. The agreement includes two five-year renewal options; if both are exercised, the total contract value could reach approximately US$16.1 billion.
Note: This content was created with AI assistance and may contain hallucinations and errors.
