Meta and DeepMind debut new models on the same day, updating open source and complex reasoning

Anthropic has released a free, open-source AI scanner designed to automatically detect vulnerabilities in open-source software. The tool periodically scans code repositories, explains potential risks, and provides code fix patches, with an official expected accuracy rate of over 90%. Open-source project maintainers can apply to use it via GitHub.

WeChat Work and WorkBuddy have released the 5.0.12 update. After upgrading, users can chat directly with WorkBuddy in the chat interface. Within permitted boundaries, the assistant can read chat messages, collaborative documents, and meeting schedules, directly invoke workbench applications to process tasks, and feed execution results back into the chat window.

Researchers tested four personal agents—Google Muse, Grok Bot, OpenAI Dots, and Manus Cue—and found that only 33% successfully and stably delegated tasks across 12 daily platforms. The tests noted that cloud-based execution still struggles to detach from local environments, and tools equipped with native machine control capabilities showed more stable performance during actual interactions.

Meta has released the open-weight model Muse Spark Open, featuring 816 billion total parameters and 42 billion active parameters. The model supports a context window of up to 1 million tokens and natively supports text, image, video, and audio input. Officials describe it as a high-performance US open-weight model, benchmarked against Qwen, GLM, and Kimi.

Google DeepMind has announced that the general-purpose model Gemini 4 Argon is officially available worldwide. The company states that the model is designed specifically to solve end-to-end, high-difficulty problems, boasting cutting-edge processing capabilities in complex coding, enterprise workflow orchestration, and cybersecurity tasks.

Nvidia has launched the open multimodal model Nemotron 3.5 Super VL, utilizing a Mixture of Experts (MoE) architecture with 120 billion total parameters and 12 billion active parameters. The model integrates complex reasoning, tool calling, and image/video/text comprehension capabilities. Nvidia states that its agent task throughput on a single GPU can reach up to 5 times that of open models of the same scale.

Google DeepMind tested four Gemini 4 Argon agents beating Minecraft without any human intervention. The agents were placed directly into a randomly generated world and, without preset task divisions or strategies, independently completed collaborative exploration, threat warnings, material collection, and error experience logging.

Anthropic reported third-quarter revenue of $16.8 billion, a quarter-over-quarter increase of about 45%, which is lower than the growth rate of over 140% in the second quarter. The company's cash flow is currently positive, generating $1.70 in revenue for every $1 spent on compute. Some investors expect its IPO valuation could be adjusted downward to under $1.5 trillion.

Alforzan, an engineering director on OpenAI's product team, was arrested by the FBI and fired by the company for allegedly installing recording equipment behind a television in a San Francisco headquarters executive meeting room. Investigations revealed the employee had queried executive schedules in internal systems and used a family member's account to ask how to hide recording devices. OpenAI's security team used a large model to analyze clues and identify the suspect.