OpenAI has released GPT-5.4 Mini and Nano, designed as fast and efficient 'subagents' for coding and other tasks. These smaller models allow a main AI model to delegate parallel tasks, significantly improving throughput and efficiency for complex operations.
Sources: 3 Covered by Latent Space, Simon Willison, The Neuron
Mistral AI has launched Forge, a platform enabling enterprises to build and train custom AI models using their proprietary data and infrastructure. This allows for the development of frontier-grade AI models tailored to specific enterprise needs while maintaining data privacy and control.
Sources: 3 Covered by Hacker News, The Neuron, The Rundown AI
NVIDIA has launched Nemotron 3 Nano 4B, a 4-billion parameter language model with a hybrid Mamba-Transformer architecture. Optimized for on-device deployment on NVIDIA Jetson and RTX GPUs, it enables faster response times, enhanced data privacy, and lower inference costs for local conversational agents.
Sources: 1 Covered by Hugging Face Blog
Chinese models now account for 41% of downloads on Hugging Face, surpassing the U.S. in open source AI model popularity. This shift reflects increasing contributions from Chinese organizations and a growing international mix of top models.
Sources: 1 Covered by Hugging Face Blog
The Omnilingual MT Team has introduced OMT, a machine translation system supporting over 1,600 languages. Utilizing specialized LLMs (OMT-LLaMA and OMT-NLLB), it achieves high-quality translation, matching or exceeding 70B LLM baselines even in low-compute settings, and includes new evaluation datasets.
Sources: 1 Covered by arXiv AI
The Qatar Computing Research Institute (QCRI) and Hamad Bin Khalifa University have launched Fanar 2.0, an Arabic-centric Generative AI platform. Developed on 256 NVIDIA H100 GPUs, it demonstrates that resource-constrained AI development can produce competitive systems with significant benchmark improvements for Arabic knowledge.
Sources: via arXiv AI
A new retrieval-augmented generation (RAG) framework powered by GPT-4 significantly accelerates the discovery of high-entropy catalysts for CO2 reduction. This method achieved 200x computational efficiency and generated highly stable and performant catalyst candidates by grounding AI creativity in physical constraints.
Sources: via arXiv AI
A large-scale red teaming competition found that 13 frontier AI models, including Claude Opus 4.5 and Gemini 2.5 Pro, are vulnerable to indirect prompt injection attacks. This highlights a critical security flaw where hidden adversarial instructions can manipulate AI agent behavior without user detection.
Sources: via arXiv AI
Researchers have discovered 'ClawWorm,' the first self-replicating worm attack targeting production-scale LLM agent frameworks like OpenClaw. This attack demonstrates autonomous infection, persistent presence, and propagation across multi-agent systems, highlighting significant security vulnerabilities in these interconnected AI environments.
Sources: via arXiv AI
A new research paper introduces SWE-QA-Pro, a benchmark for agentic repository-level code understanding, and a scalable training recipe. This approach allowed a Qwen3-8B model to surpass GPT-4o on the benchmark, demonstrating a significant advancement in training smaller models for complex software engineering tasks.
Sources: via arXiv AI
Researchers introduce Hilbert, an agentic framework that combines informal LLM reasoning with formal verification to recursively build mathematical proofs. Hilbert significantly outperforms existing methods on benchmarks like miniF2F and PutnamBench, setting a new public benchmark.
Sources: via arXiv AI
Researchers developed an automated framework utilizing Large Language Models and a novel symbolic algorithm to unify mathematical formulas. Applied to 455,050 arXiv papers, the system validated and proved relations between hundreds of formulas for constants such as pi, e, and Catalan's constant.
Sources: via arXiv AI
Google is expanding its 'Personal Intelligence' feature across its AI Mode in Search, the Gemini app, and Gemini in Chrome for U.S. users. This feature enables highly tailored responses by securely connecting user data from various Google applications like Gmail and Photos.
Sources: via Google AI Blog
Google, Amazon, Anthropic, Microsoft/GitHub, and OpenAI have collectively committed $12.5 million to the Linux Foundation’s Alpha-Omega Project and OpenSSF. This funding aims to enhance the security of open source software by deploying fixes and providing advanced AI tools to maintainers, addressing a new generation of AI-driven threats.
Sources: via Google AI Blog
Google DeepMind has introduced "Measuring Progress Toward AGI: A Cognitive Taxonomy," a new paper outlining a scientific framework based on cognitive science to evaluate AI systems' general intelligence across 10 key cognitive abilities. This framework aims to provide empirical tools to track and understand advancements toward Artificial General Intelligence.
Sources: via DeepMind Blog