VibeThinker-3B, a 3 billion parameter model, has been introduced as a compact dense model for enhanced verifiable reasoning.

The model leverages a Spectrum-to-Signal post-training paradigm with curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation. It scores 94.3 on AIME26, improving to 97.1 with test-time scaling, and 80.2 Pass@1 on LiveCodeBench v6. Additionally, it achieves a 96.1% acceptance rate on unseen LeetCode contests and scores 93.4 on IFEval. These results suggest that compact models can match or exceed larger flagship models like DeepSeek V3.2 and GLM-5.

Mistral released OCR 4, an enterprise document AI model supporting 170 languages with bounding boxes and self-hosted deployment.

This model includes features like bounding boxes, block classification, and inline confidence scores, enhancing text extraction with context. OCR 4 outperformed other systems in tests, achieving a 72% win rate in preferences and an 85.

F3 is an open-source data file format focused on efficiency, interoperability, and extensibility.

F3 addresses layout issues found in older formats like Parquet, using embedded Wasm decoders for future-proofing. This project is a research prototype and not intended for production use.

A logging bug in Codex's CLI caused excessive data writing to local SSDs, potentially reaching 640 TB per year.

The issue stemmed from Codex SQLite feedback logs being the main source of continuous data writing. On a 1 TB SSD, this could result in approximately 640 full-drive writes annually, surpassing typical consumer SSD endurance ratings of around 600 TBW. Two pull requests have been merged to address the problem, reportedly reducing log volume by 85% and closing the issue.

OpenAI introduced new Daybreak tools, including Codex Security and GPT-5.5-Cyber, for organizational vulnerability management.

These tools assist in finding, validating, and patching vulnerabilities at scale. Codex Security is designed to streamline the identification process, while GPT-5.

Researchers highlight that human brains, evolved to focus on threats, are now overwhelmed by constant bad news from around the world.

Despite the instinct to stay informed, many people report avoiding news to manage their mood and feelings of helplessness. According to a 2025 Reuters Institute report, 69% of Canadians and 40% globally sometimes avoid the news. This behavior reflects a natural response to an environment our brains weren't designed to handle, emphasizing the need for healthier news consumption habits.

GPT-5.5 is reported to hallucinate three times more than the MIT-licensed GLM-5.2 in AI model evaluations.

GPT-5. 5, with an estimated 1-2 trillion parameters, scored an 86% hallucination rate on the AA-Omniscience benchmark.

GitHub developed Qubot, an internal analytics agent powered by Copilot, to enable employees to query data using plain language.

Qubot allows GitHub staff to interact with data without requiring specialized query languages or tools. The project was led by Matteo Vasirani, who manages product analytics and data science at GitHub. This initiative highlights the integration of AI to streamline internal data processes.

The Enterprise-Managed Authorization extension for the Model Context Protocol has reached stable status, allowing centralized MCP server access through an organization's identity provider.

With this extension, users can access all connected MCP servers on their first login without needing per-app OAuth configurations. Organizations like Anthropic, Microsoft, and Okta are adopting this approach. Administrators set access policies once, and the identity provider manages access decisions based on user roles and group memberships, streamlining the process significantly.