AI Daily | Meta Releases Muse Spark Open, Microsoft Details MAI-Base-2, and Apple Intelligence China Plans Emerge

Model Releases & Updates

Muse Spark Open MMeta

• Bottom line：Meta released Muse Spark Open, an 816B parameter open-weight Mixture-of-Experts multimodal model featuring a 1-million-token context window.
• Architecture：Features 816 billion total parameters with 42 billion active parameters per token in an open-weight MoE framework.
• Capabilities：Natively multimodal across text, image, video, and audio, specifically optimized for long-horizon coding and software engineering.
• Availability：Available via Hugging Face under the Muse Model License with day-zero support from major inference providers.
• Source：PyTorch Blog

MAI-Base-2 MMicrosoft AI

• Bottom line：Microsoft AI detailed the pre-training of MAI-Base-2, a 2.8-trillion parameter MoE base model boasting high FLOPs efficiency.
• Scale & Efficiency：Combines 2.8 trillion total parameters with 80 billion active parameters, achieving up to 12.3x higher training FLOPs efficiency than baseline models.
• Performance：Achieved lower bits-per-byte scores on held-out code benchmarks compared to existing open-weight equivalents.
• Source：Microsoft AI News

Nemotron 3.5 Super VL NNVIDIA

• Bottom line：NVIDIA unveiled Nemotron 3.5 Super VL, a 120B parameter open multimodal model designed for long-running agent workflows.
• Parameters：Contains 120 billion total parameters with 12 billion active parameters per token.
• Performance：Demonstrated an 87% pass rate on behavioral agent suites while significantly reducing task completion times.
• Source：NVIDIA Blog

Product Releases & Updates

ChatGPT Chrome Extension Browser Tab Control OOpenAI

• Bottom line：OpenAI upgraded its ChatGPT Chrome extension to allow users to inspect and operate active browser tabs using conversational commands.
• Functionality：Enables ChatGPT to view opened tabs, compare data across multiple pages, fill out online forms, and organize browsing sessions.
• Integration：Accessible via the web or desktop application by referencing @Chrome in conversational prompts.
• Source：The Verge Coverage

Claude Code Review Generally Available AAnthropic

• Bottom line：Anthropic made Code Review generally available for Team and Enterprise customers, offering automated PR audits with adjustable depth settings.
• Performance：Features review speeds up to three times faster than the research preview version with enhanced bug detection.
• Customization：Administrators can select between Quick, Standard, and Deep review levels for individual repositories.
• Source：Claude Blog

Code Layers and Payload Integration FFigma

• Bottom line：Figma announced the open beta of Code Layers and the integration of Payload CMS to bridge design and code directly on the canvas.
• Code Layers：Allows teams to explore and build code-backed prototypes natively within the Figma canvas.
• Payload Integration：Automatically provisions databases, user authentication, and admin panels when building data-driven web applications.
• Source：Figma Blog

Industry News

Anthropic Reports Q3 Revenue and IPO Outlook AAnthropic

• Bottom line：Anthropic disclosed a Q3 revenue of $16.8 billion to investors ahead of its planned November IPO, while noting decelerating growth.
• Financial Metrics：Third-quarter revenue reached $16.8 billion, representing a 45% increase from the previous quarter.
• Efficiency：Reported positive cash flow, generating $1.70 in revenue for every dollar spent on computing power.
• Source：Simon Willison Report

Apple Intelligence Planned for China Launch in December AApple

• Bottom line：Reports indicate Apple will launch Apple Intelligence in mainland China in December with iOS 27.3, utilizing localization partners.
• Initial Features：Will include foundational features such as notification summaries, writing tools, and photo clean-up utilities.
• Partnerships：Cloud model capabilities in China will be provided by Alibaba, with Baidu handling local search and Siri adaptations.
• Source：MacRumors Coverage

Research Papers

Magic State Cultivation in Quantum Computing GGoogle Quantum AI

• Bottom line：Google demonstrated successful magic state cultivation on quantum hardware, achieving a 40x error reduction on the Willow processor.
• Publication：Published in the journal Science, validating error-correction techniques on physical quantum hardware.
• Results：Achieved a 40x error reduction and attained state-of-the-art 99.99% logical fidelity compared to previous methods.
• Source：Google Research Blog

ExtractBenchPro LLlamaIndex

• Bottom line：LlamaIndex introduced ExtractBenchPro, a rigorous benchmark designed to evaluate complex real-world document extraction tasks.
• Scope：Curates complex long-tail document scenarios including trial balances, well logs, and financial reconciliations.
• Objective：Addresses the saturation of existing extraction benchmarks by testing advanced multi-step reasoning and internal jargon cross-referencing.
• Source：

  View post on X by @jerryjliu0

  在 X 上查看 @jerryjliu0 的貼文 ↗

Other Highlights

Perplexity Open Sources SPACE Sandbox Platform PPerplexity

• Bottom line：Perplexity open-sourced SPACE, the secure microVM sandbox platform that powers the execution environment of Perplexity Computer.
• Architecture：Assigns an isolated Firecracker microVM to each agent task, controlling credentials and outbound traffic externally.
• Open Source：Released under the Apache 2.0 license on GitHub to raise security standards across agentic platforms.
• Source：Perplexity GitHub Repository