XREAL has launched AURA, the first wired XR glasses powered by Android XR, featuring integrated Gemini AI for real-time assistance. The device uses a separate compute puck to maintain lightweight form factors while projecting apps into the user's field of view, with pre-orders opening in the US, Canada, and Korea.
Multimodal LLM
多模态大模型
Deep learning models capable of processing multiple data types like text and images simultaneously
Curated Intelligence
A team from Greater Bay Area University used a Chinese multimodal large language model to solve the 3D Ghys conjecture, a major mathematical problem that had persisted for over 30 years. The proof was formally verified by Peking University's SHEAF framework. This marks the first time Chinese scientists have leveraged a domestic LLM to crack a world-class math problem, validating a new research paradigm combining expert intuition with machine-assisted solving.
A comparative study reveals that a fine-tuned Inception v3 CNN (2015) outperforms frontier multimodal LLMs like GPT-6 and Claude in skin cancer diagnosis. On a dataset of 200 biopsy-proven lesions, the LLMs misclassified malignant melanoma as benign seborrhoeic keratosis, recommending discharge, while the specialized CNN correctly identified the malignancy. This highlights the limitations of general-purpose vision-language models compared to domain-specific small models in critical vertical medical tasks.
Scale AI and Elorian introduced the 'Humanity's Sixth Sense' (HSS) benchmark to evaluate multimodal LLMs on intuitive reasoning in implicit contexts. Results show humans averaging 93.1%, while the top model, GPT-6-astra, scored only 53.6%, with a median of 30.9%. This highlights a critical gap in models' ability to grasp physical and social nuances without explicit cues.
NVIDIA's latest open-source multimodal model, Nemotron 3.5 Super VL, is now available for free on OpenCode. Featuring a Mixture-of-Experts architecture with 120B total and 12B active parameters, it is optimized for long-running agent tasks.
Bland has launched iMessage-based AI agents that allow businesses to generate branded customer service bots in seconds using only a website URL. These multimodal agents can parse images like insurance cards and documents, enabling end-to-end workflows such as booking appointments, processing payments, and sending calendar invites directly within the chat thread.
Nvidia has released and open-sourced Nemotron 3.5 Super VL, a 120-billion parameter Mixture-of-Experts multimodal model designed to power continuously running enterprise AI agents. By activating only 12 billion parameters per token, it efficiently processes text, images, documents, and videos. Alongside this flagship release, Nvidia introduced two compact Bolt models targeting local PCs and edge devices, expanding the deployment capabilities for on-device intelligent agents.
NVIDIA's Nemotron 3.5 Super VL is now live on OpenRouter. This open-source 120B-A12B multimodal model is designed for long-running agents, capable of reasoning across text, images, documents, GUIs, and video with a 1 million token context window.
A team from Shanghai AI Lab solved the 30-year-old 3D Ghys conjecture using the S2 multimodal scientific model via a human-AI collaborative approach. The proof was formally verified by Peking University's SHEAF framework, confirming its correctness. This breakthrough validates a new research paradigm integrating expert intuition into machine reasoning loops, with papers and code now publicly available.
Core specs for Samsung's upcoming first smart glasses have leaked, featuring a ~41g weight and the Snapdragon AR1 Gen1 chip. The display-less device relies on open-ear speakers and a quad-mic array for Gemini voice interactions, with a launch expected next month.