Search AI Products and News

VLM2Vec-V2, a multimodal embedding framework developed by Salesforce Research and other institutions, breaks through the limitations of traditional models. Based on the Qwen2-VL architecture, this framework innovatively unifies image, video, and visual document retrieval tasks. It adds five new evaluation tasks to expand the MMEB dataset. With key technologies such as dynamic resolution and M-RoPE, it achieved an average score of 58.0 across 78 datasets, leading in performance, especially in video tasks. Although its document retrieval is slightly behind ColPali, it

Apple's MLX framework now supports CUDA, marking a strategic shift to embrace NVIDIA's ecosystem. This enables developers to train models on NVIDIA GPUs and deploy on Apple devices, enhancing efficiency and bridging platform gaps.....

Open-source RL framework ART released, enhancing AI Agent training with GRPO tech. Features Python support, small model compatibility, and client-server architecture for multi-step tasks like email automation and game AI.....

Recently, Kunlun Tech Skywork and Nanyang Technological University jointly launched an innovative hierarchical multi-agent collaboration framework called AgentOrchestra, which is referred to as the 'AI Symphony Orchestra,' marking a significant step forward in agent collaboration within the field of general artificial intelligence. Faced with complex tasks in the real world that challenge single models or agents, AgentOrchestra enables intelligent agents with distinct expertise to collaborate and tackle challenges together, by mimicking the collaborative model of a symphony orchestra. Traditionally, general AI

Apple's MLX framework now supports CUDA, allowing developers to create applications on Apple Silicon devices and export them to run in a CUDA environment. This move significantly reduces development costs, as small teams no longer need to purchase expensive Nvidia hardware upfront. The project, led by GitHub developers, has been integrated into the main MLX branch. Although it does not support running directly on Macs with Nvidia GPUs, development and testing can be fully completed on Apple devices. Performance in CUDA environments is significantly better than on Macs, and combining it with Nvidia hardware for AI applications is highly beneficial.

ByteDance's Pico is developing a lightweight MR glasses, weighing only 127 grams. It uses a split design to transfer computing tasks to an external processing unit and is developing a dedicated chip to reduce latency. This move directly targets Meta's MR strategy, as the latter has paused the development of Quest 4 and shifted towards lightweight devices. Both major companies are betting on split-style MR glasses, indicating that the industry is transitioning from traditional VR headsets to more compact forms.

Hong Kong Polytechnic University and OPPO Research Institute jointly released the open-source video super-resolution framework DLoRAL, which generates high-definition videos in one step using diffusion models. The framework adopts a dual LoRA architecture: C-LoRA maintains temporal consistency between frames, while D-LoRA enhances spatial details. Through a two-stage training strategy, it optimizes temporal coherence and high-frequency information. Compared to traditional methods, DLoRAL improves inference speed by 10 times while maintaining smoothness, significantly enhancing image details, and providing an efficient open-source solution for video HD enhancement.

Welcome to AIbase's [AI Daily Report]! Spend three minutes every day to learn about the latest AI news, helping you understand AI industry trends and innovative AI product applications. For more AI updates, visit: https://www.aibase.com/zh1. Tencent open-sources the lightweight Huyuan-A13B model, which can be deployed with just one mid-range GPU card. Tencent has released a new member of the Huyuan large model family, the Huyuan-A13B model, which uses a mixture of experts (MoE) architecture, with a total parameter scale of 80 billion and an activated parameter count of 13 billion, large

Tencent officially launched and open-sourced a new member of the Hypermix large model family - the Hypermix-A13B model. The model adopts an expert mixture (MoE) architecture, with a total parameter scale of 80 billion and an activated parameter count of 13 billion. It maintains the performance of top-tier open-source models while significantly reducing inference latency and computational costs, providing a more cost-effective AI solution for individual developers and small and medium-sized enterprises.

On June 18th, the artificial intelligence research institute OpenAI officially open-sourced its customer service agent framework, marking an important step in the company's layout of enterprise-level AI solutions. This framework aims to promote the practical commercial application of intelligent agent systems from laboratory environments to real-world scenarios by providing transparent toolchains and clear implementation examples.

At the 2025 WWDC conference, Apple announced the opening of its artificial intelligence technology to the developer community, releasing the Foundation Models framework and upgraded Xcode26, marking a major breakthrough for Apple in the field of AI development tools. Foundation Models Framework: Privacy-Focused AI Integration Solution The new Apple Foundation Models framework provides developers with a centralized set of tools to directly embed Apple's self-developed AI models into applications.