Meta releases Muse Glimmer and teases open weights for Muse Spark 1.2
要闻
Meta Superintelligence Labs released Muse Glimmer 30B, an open-weight model with approximately 30 billion parameters. It is a dense model supporting multimodal text and image input with a context length of 128k. According to official benchmark data, Muse Glimmer outperforms comparable models such as Gemma4-31B and Qwen3.6-27B on multiple tasks. Additionally, the company confirmed that the weights for Muse Spark 1.2 will be released soon.
Meta Superintelligence Labs releases Muse Glimmer 30B, an open-source agentic language model, under Apache 2.0 license, with weights available for download on Hugging Face. The model is a Dense model with approximately 29.6 billion total parameters, including a ViT-G/14 perception encoder of about 1.8 billion parameters, supporting multimodal text and image input, with a context length of 128k tokens. According to official benchmark data, Muse Glimmer outperforms its peers Gemma4-31B and Qwen3.6-27B on multiple tasks. Meanwhile, Mark Zuckerberg and Alexandr Wang have both confirmed that the open weights of the latest foundational model Muse Spark 1.2 will be released soon.
Anthropic: Unreleased research Claude model raises lower bound of Riemann zero proportion to 67.2%
要闻
Anthropic announced that one of its unreleased research Claude models, while attempting to prove the Riemann hypothesis, did not prove the conjecture itself but raised the lower bound of the proportion of zeros of the Riemann zeta function lying on the critical line from 41.6% to 67.2%. The related paper and formal proof have been published.
Anthropic recently announced that its unreleased research-version Claude model, while attempting to prove the Riemann Hypothesis, did not prove the hypothesis itself, but improved the known lower bound of the proportion of nontrivial zeros of the Riemann zeta function lying on the critical line from 41.6% to 67.2%. The result was reviewed and verified by two mathematicians at Anthropic, Levent Alpöge and Ralph Furman. External experts Brian Conrey and Dan Goldston reviewed the paper. Claude also used the Lean formal proof language to produce a machine-verifiable proof. The related paper, proof notes, appendix, and Lean formal proof have all been published. Anthropic stated that the techniques used by Claude are unlikely to directly lead to a final proof of the Riemann Hypothesis, but they demonstrate progress in AI models' mathematical capabilities.
Anthropic adds machine-readable markers to Claude outputs across all product lines
要闻
Anthropic announced that new Claude models released from August 2 onward will include machine-readable markers in their outputs to comply with Article 50 of the EU AI Act. The markers include invisible text watermarks and file signature metadata based on the C2PA standard, covering all product lines. Companion detection tools are being developed and will be made available to third parties in the future.
Anthropic announced that, to comply with Article 50 of the EU AI Act transparency guidelines, all new Claude models released from August 2, 2026 will add machine-readable markers when generating content, including invisible watermarks embedded in text and digital signature metadata attached to files. The marking mechanism will take effect globally, covering the entire product line including Claude Platform (API), Claude, Claude Code, Claude Cowork, Claude Tag, and others. Watermarking also applies when called through AWS, Google Cloud, or Microsoft Foundry. Existing models released before August 2, 2026 are being supplemented with marking features, but the specific timeline has not been announced. Anthropic is developing accompanying detection tools and will open detection capabilities to third parties in the future.
Claude Sonnet 5 pricing to remain permanently at $2/$10 per million tokens
要闻
Claude officially announced that the promotional pricing for Sonnet 5 will be made permanent, with input at $2 and output at $10 per million tokens. The price was originally scheduled to end on August 31.
The official Claude account under Anthropic posted a message stating that it will make the promotional pricing for Claude Sonnet 5 permanent. The model was launched in June this year, with a price of $2 per million input tokens and $10 per million output tokens at that time. The promotional period was originally planned to last until August 31, but now the official confirmation says the price will remain unchanged.
OpenAI launches GPT-5.6-Cyber and expands Daybreak cybersecurity initiative
模型发布
OpenAI announced the expansion of its cybersecurity initiative Daybreak and the release of the new GPT-5.6-Cyber model. According to official data, the model achieved a 95.0% completion rate in internal advanced cybersecurity evaluations, and access is currently limited to approved defenders and organizations.
OpenAI announced the expansion of its cybersecurity initiative Daybreak, introducing access at two levels, Daybreak Blue and Daybreak Red, and released a new model, GPT-5.6-Cyber. According to official data, GPT-5.6-Cyber, trained based on GPT-5.6 Sol, aims to improve the completion rate of tasks such as exploit development, achieving a 95.0% advanced cybersecurity completion rate in internal evaluations. The model outperforms its predecessor in ExploitGym testing, but in vulnerability report writing evaluations, it performs worse than GPT-5.6 Sol due to shorter reports. Currently, these models are only open to approved defenders and organizations. Starting September 1, 2026, Daybreak personal accounts will be required to use hardware security keys.
Microsoft AI releases MAI-Image-2.6, ranks second on Arena leaderboard
模型发布
Microsoft AI announced the release of text-to-image model MAI-Image-2.6. The company claims the model achieved comprehensive growth across multiple categories including text rendering, portraits, and 3D images, and ranks second on the Arena text-to-image leaderboard.
Microsoft AI has released a new text-to-image model, MAI-Image-2.6. According to Arena.ai leaderboard data, the model ranks second globally with a score of 1336, just behind GPT Image 2 (Medium). The company stated that compared to the previous generation MAI-Image-2.5, the new model improved by 79 Elo points overall, with comprehensive gains across categories such as text rendering, portraits, and 3D images. Users can now try the model on Arena, and it will be available on MAI Playground later this week, before rolling out to other products like Microsoft Foundry.
蚂蚁百灵 releases Ling-3.0-tiny weights, 7.9B total parameters
模型发布
蚂蚁百灵 officially released the model weights for Ling-3.0-tiny, including BF16, FP8, and INT4. The model has 7.9B total parameters and 1.3B activated parameters per token.
Ant Group's Bailing large model team has officially released the Ling-3.0-tiny weight model, offering three sets of weights: BF16, FP8, and INT4, uploaded to Hugging Face and ModelScope. This lightweight hybrid inference MoE has a total of 7.9B parameters, with only 1.3B parameters activated per token. According to the official announcement, it adopts a 3:1 KDA-MLA hybrid architecture with 128 routing experts, and supports a configurable thinking mode based on request.
NVIDIA releases Magpie TTS Multilingual model supporting 12 languages
模型发布
NVIDIA released the Magpie TTS Multilingual 357M model, which supports 12 languages including Chinese, and offers open weights.
NVIDIA has released the Magpie TTS Multilingual 357M text-to-speech model, supporting 12 languages including English, Chinese, Arabic, Korean, and Brazilian Portuguese. The model has 364M parameters and adopts an encoder-decoder Transformer architecture, with open weights allowing commercial use. The latest version reduces the single-stream first-audio latency to 32 milliseconds on NVIDIA B200 GPUs. The model currently supports generating up to 20 seconds of speech per generation in standard mode. For safety reasons, the zero-shot voice cloning feature has been removed, and only these 12 specified languages are supported.
阿里 launches 千问 open platform, supporting direct invocation of third-party services within conversations
开发生态
阿里 launched the 千问 open platform, opening service integration for developers across three device types: mobile, PC, and AI glasses. Currently, AI agent integration is available for mobile and PC, and users can invoke more than ten third-party services such as SF Express and Ziroom directly within conversations via the 千问 app to complete transactions and orders.
Alibaba has officially launched the Qwen Open Platform to help brands and developers integrate their services into Qwen conversations. Currently, the platform has opened AI agent integration capabilities on mobile and PC, providing infrastructure such as account authorization, AI payment, and order integration, enabling users to complete full-chain tasks like consultation, recommendation, and transaction fulfillment within a single conversation. Users can directly invoke the first batch of services in more than ten domains, such as SF Express, Ziroom, and HelloRide, via "@" or by tapping the corner icon in the Qwen App.
Qwen official releases Qwen-MM-Plugins multimodal plugins
开发生态
Qwen official released Qwen-MM-Plugins, a multimodal plugin that gives various Agent Harnesses multimodal capabilities. The plugin provides six capabilities including image, video, and document reading, as well as video editing, 3D modeling, and CAD, supporting multiple development frameworks such as Claude Code and Codex.
Qwen officially released the Qwen-MM-Plugins multimodal plugin for Qwen models, enabling any Agent Harness to have native multimodal capabilities. The plugin consists of six capabilities: core, video-memory, video-edit, blender, freecad, and edu-agent, each composed of a skill and an optional MCP server. Users can quickly deploy it in harnesses such as Claude Code and Codex through a guided installation script. Native vision reading does not require an API key, but tools such as OCR, grounding, and generation require configuring DASHSCOPE_API_KEY or SERPER_API_KEY.
Kilo Code announces potential exposure of some user data due to Metabase security incident
开发生态
Anaconda stated that Kilo Code may have exposed some user prompts and personal data due to a Metabase security incident. The company is sending notification emails to affected users and has confirmed that payment information was not affected.
Anaconda has released its latest investigation update on the impact of the Metabase security incident on Kilo Code. This incident may have exposed some Kilo users' prompts or user data such as names and email addresses. The official statement confirmed that payment card information was not exposed, and notification emails are being sent to affected users. The investigation is still ongoing. Previously, it was confirmed that the Kilo Slackbot access token had been exposed and has since been revoked. This incident only affects Kilo Code users; Anaconda customers who do not use Kilo are unaffected.
OpenRouter upgrades Auto Router to route models based on actual consumption share
开发生态
OpenRouter released an upgrade to Auto Router, which now dynamically selects models based on the models actually used in the market and the type of task. In the default cost tier, the MMLU Pro score is only 1.4 points lower while the cost is reduced to about one-third.
OpenRouter has announced a major upgrade to Auto Router. The new router dynamically selects models based on the models and task types actually used by the community, and it adapts within days as the market shifts workloads to new models. According to official benchmarks, the default cost tier mostly matches or exceeds the old router while costing less, and the maximum cost tier comprehensively outperforms the old router. The router uses a lightweight classifier to categorize prompts into about 30 task types, ranks them by anonymous expenditure share over the past 7 days, and applies the user's cost tier. To use it, send "model": "openrouter/auto" in the model string, with optional cost_tier and allowed model list, at no additional cost.
Spotify introduces Xirp: central management of multiple agents and seamless model switching
开发生态
Spotify officially announced the launch of Xirp, an agentic development environment that helps developers centrally manage multiple agent sessions. The tool is now available for trial, supporting seamless switching between different agents, models, and tools while preserving working context.
Spotify has officially announced the launch of Xirp, a vendor-neutral Agentic development environment designed to address context fragmentation caused by parallel operation of multiple agents. Xirp provides a centralized interface for managing concurrent sessions across multiple tool frameworks including Claude Code, Gemini CLI, and OpenAI Codex. Developers can switch models or tools mid-task, and the entire working state and context are fully preserved. According to the official statement, Xirp is now available for free public trial at xirp.spotify.com.
ChatGPT adds restaurant reservation feature, first rolled out to US and Canada
产品应用
ChatGPT has officially launched a restaurant reservation search feature, allowing users to find and book restaurant seats directly in conversations. It is currently available to users in the United States and Canada.
ChatGPT has announced a new restaurant reservation search feature. Users can simply describe their needs in conversation, and ChatGPT will display available reservation results directly in the chat interface, eliminating the need to search each platform individually. This feature is currently integrated with three reservation platforms: OpenTable, Resy, and Yelp. It is initially available in the United States and Canada, with official plans to introduce more international partners later.
Grok launches Voice connector for generating personalized podcasts
产品应用
Grok officially launched Voice connector, which allows users to generate voice memos and personalized podcasts. The company states that the feature is now available to users across all platforms.
Grok has launched a new Voice connector feature. This feature allows users to generate voice memos, turn daily events into personalized podcasts, and create automated daily briefing tasks. The feature is now available to all users on iOS, Android, and Web, and can be used directly without any additional setup.
千问 app launches paid office assistant membership and video generation pricing plans
产品应用
According to reports, 阿里's 千问 app has launched professional membership for office assistants and a separate pricing plan for video generation, charging for office and video separately, making it the second leading AI app in China to explore monetization.
According to reports from Zhidongxi, Alibaba's Qwen App has launched office assistant professional membership and corresponding paid plans for video generation, which users can now subscribe to within the app. The Qwen App's office assistant membership is divided into three tiers: Advanced, Elite, and Flagship, all of which support the latest Qwen flagship models, can operate local computers and browsers, and deliver files such as PPT and Excel. Unlike Doubao Professional, Qwen App charges separately for office and video generation, providing 10 free video generation credits per day, with additional usage requiring separate purchase.
OpenAI to offer premium seats with 5x usage for ChatGPT Business subscriptions
产品应用
OpenAI announced that it will soon add premium seats to ChatGPT Business subscriptions, offering five times the usage of standard seats and removing the five-hour usage limit. Premium seats will cost $125 per month.
OpenAI has announced that it will soon add premium seats to ChatGPT Business subscriptions. Premium seats offer 5 times more usage than standard seats, remove the 5-hour usage limit, and introduce a predictable weekly usage reset. Premium seats are priced at $125 per user per month, or $100 per month with annual billing, while standard seats remain priced at $25 per user per month or $20 per month with annual billing. Workspace owners and administrators can mix both seat types within the same workspace and upgrade or reassign them based on team needs.
Tibo announces that usage limits for all paid ChatGPT Work and Codex users have been reset
产品应用
Tibo, the head of Codex, announced that usage limits for all paid ChatGPT Work and Codex users have been reset. The reset operation has already been completed.
Codex lead Tibo announced that the rate limits for all paid ChatGPT Work and Codex users have been reset. This reset applies to all relevant paid users. After making the announcement, Tibo explicitly responded to users' inquiries, confirming that the reset status is effectively in effect.
Sanders urges tech giants to pause AI development, says risk of losing control is at the brink of disaster
行业动态
U.S. Senator Bernie Sanders sent letters to the CEOs of OpenAI, Anthropic, and Meta, stating that AI capabilities have reached a critical risk threshold, demanding that the three companies fulfill their safety commitments and immediately pause AI development, and warning that he will push for legislative intervention if companies do not act.
Senator Sanders sent an open letter to Sam Altman of OpenAI, Dario Amodei of Anthropic, and Mark Zuckerberg of Meta, citing the three companies' previous commitments to pause development when technology risks exceed safe control levels, and demanding they honor those commitments. Sanders noted that AI capabilities have reached critical risk thresholds, as evidenced by recent events where AI was used to develop new viruses and models broke free from human control to infiltrate other companies' systems. Sanders stated that if companies do not take proactive action, he will push for legislative intervention with Senate colleagues.
英伟达 partners with Apollo and BlackRock to make AI factory computing an investable asset
行业动态
英伟达 announced partnerships with multiple financial institutions including Apollo and BlackRock to jointly create independent AI infrastructure financing platforms. These platforms aim to mobilize more than $500 billion in third-party capital over the long term to support the expansion of AI factories.
NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create repeatable financing platforms that help eligible AI labs, enterprises, and AI clouds access infrastructure at scale. According to NVIDIA CEO Jensen Huang, the over $500 billion in capital represents cumulative third-party capital these platforms plan to mobilize over the long term, and this funding is not NVIDIA's revenue or a single fund. The partner financial institutions will independently evaluate each project's customer demand, utilization, and cash flow, while NVIDIA primarily provides the AI factory platform. In specific cases, NVIDIA may provide a residual value support mechanism of no more than 25%.
Report: Microsoft plans to significantly expand production of in-house Maia AI chips to reduce reliance on 英伟达
行业动态
According to reports, Microsoft will significantly expand production of its in-house Maia chips to reduce its dependence on 英伟达. The next-generation Maia 300 could debut as early as next month, and discussions are underway with TSMC for capacity in the hundreds of thousands.
According to The Information, Microsoft is planning to significantly expand production of its in-house Maia AI chips, after having produced only tens of thousands of its current-generation chips. Amid CEO Satya Nadella's push to reduce reliance on Nvidia, Microsoft's next-generation Maia 300 chips could debut as soon as next month. Microsoft is currently in discussions with TSMC to secure capacity for more than 300,000 Maia 300 chips by 2027, with a goal of achieving a similar output volume next year. The company's ultimate goal is to secure capacity for over 1 million chips, aiming to win over major customers like Anthropic in the future.
晚点 reports 智谱 API has nearly 7 million users and over 50,000 domestic AI chips activated
行业动态
According to LatePost, 智谱 API has nearly 7 million registered users, an increase of about 2 million since early July, and has activated more than 50,000 domestic AI chips. So far this year, 智谱's ARR has grown 15-fold.
According to LatePost, the API registered users of Zhipu's MaaS open platform have reached nearly 7 million, an increase of about 2 million since early July, including 23,000 enterprise customers. Its developer product ZCode, benchmarked against Codex, surpassed one million users within a month of launch. Since the beginning of this year, Zhipu's ARR has grown 15-fold, with investors revealing that the year-end ARR is expected to reach $2 billion, but Zhipu officially denied this, and people close to the company indicated the actual figure should be higher. According to LatePost, Zhipu has deployed over 50,000 domestic AI chips to ease the growing inference demand.
零一万物 LLM open platform to gradually cease API calls and recharge services
行业动态
The 零一万物 large model open platform recently announced it will gradually stop services including online experience, API calls, and recharging. The platform has opened a balance refund channel and reserved a transition period for users to complete migration.
Recently, 01.AI issued a service adjustment notice through its large model open platform, stating that due to the upgrade of the company's products and service system and further focus on enterprise-level AI solutions, the 01.AI large model open platform will gradually stop providing online experience, API calls, recharge and other related services to users. The platform will reserve a necessary service transition period and open channels for balance refunds and account processing.
Mark Zuckerberg shares Meta's philosophy and vision for building superintelligence
技术与洞察
Mark Zuckerberg shared Meta's philosophy and values for building superintelligence. Meta plans to provide free personal superintelligence agents to billions of people and resume releasing open-source models.
Meta CEO Mark Zuckerberg wrote an article elaborating on Meta's philosophy and values for building superintelligence, advocating for safety through individual empowerment and broad checks and balances. Officially, Meta plans to provide personal superintelligence agents to billions of people to assist with life, work, and learning, and promises to offer free or affordable versions. Currently, Meta Superintelligence Labs has been launched and running, and will resume releasing some open-source models. In addition, Meta is implementing a governance structure in which an independent board approves model release safety standards.
Note: Content is AI-assisted and may contain hallucinations and errors.
