Ai

Technology

• Researchers train Iris search agents using web hyperlink structures: Researchers have introduced Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales. The models were trained using a data pipeline that reverse-constructs multi-hop question chains from the hyperlink structure of a web corpus.

  AI Models • arXiv

  Permalink

• NVIDIA Doubles Down on Local AI With PAIR: NVIDIA has announced the launch of the Personal AI Router (PAIR), a new form of local router designed to enhance local AI capabilities. The open-source project allows users to route and manage AI workloads locally on RTX-powered hardware.

  NVIDIA • Sam Witteveen

  Permalink

• Saudi firm HUMAIN builds Arabic AI model on MiniMax open stack: The PIF-backed company HUMAIN has launched an Arabic frontier AI model built on MiniMax's open Mixture-of-Experts stack. This development highlights how sovereign AI projects are increasingly utilizing Chinese open-weight models as foundational bases for national initiatives.

  AI Models • Pandaily

  Permalink

• Alibaba's Qwen team releases open-source autonomous driving model: Alibaba's Qwen team, in collaboration with Huazhong University of Science and Technology, has released Qwen-Drive-1.0-4B, an open-source model for autonomous driving. The model integrates driving-scene understanding with trajectory planning, utilizing Qwen3.5-4B as its vision-language foundation.

  AI Releases • TechNode

  Permalink

• Fields Medalist helps build hybrid mobile-cloud AI system for ARC-AGI: A new hybrid AI system combining a local 4B Qwen model on a mobile device with a cloud-based GLM model has been developed to tackle the ARC-AGI benchmark. The project involves collaboration with a Fields Medalist.

  AI Models • 量子位

  Permalink

• New inference-time method reduces hallucinations in OpenAI's Whisper: Researchers have developed a training-free, inference-time method to reduce hallucinated transcripts in OpenAI's Whisper model. The approach estimates a compact hallucination-associated subspace from non-speech calibration data and projects decoder hidden states away from it.

  AI Research • arXiv

  Permalink

• DeepSeek plans 160,000-chip Huawei AI cluster in Inner Mongolia: DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a new data center in Inner Mongolia. The proposed installation would be among the largest known clusters of Huawei AI chips. However, the deployment schedule remains dependent on Huawei's production capacity.

  Hardware • TechNode

  Permalink

• More capable LLMs can increase systemic risk in financial markets: A study shows that improving individual large language model capability can degrade system-level outcomes rather than improve them. Researchers hypothesize that shared training and architectures cause more capable LLMs to behave similarly, creating correlated actions that do not diversify away.

  Research • arXiv

  Permalink