ByteDance's Volcano Engine reports a 137-fold surge in daily usage of its Doubao AI model, now processing 1.64 trillion tokens daily. The company has launched new products including image editing and interpretation models, while maintaining a dominant 46.4% market share in China's public cloud AI services.
DeepSeek's groundbreaking Native Sparse Attention (NSA) technology, developed with Peking University, won the ACL 2025 Best Paper Award. The innovation achieves 11x faster long-text processing while outperforming traditional models, extending context length to 1 million tokens through dynamic hierarchical sparsity and parallel attention branches.
Midjourney has launched a 'Recommended for You' feature, offering users AI-generated images and videos tailored to their preferences. The one-click tool analyzes user interactions to deliver customized content, enhancing productivity and personalization. The feature supports style and chaos parameters for further refinement.
Alibaba's Tongyi Lab has released WebShaper, a groundbreaking open-source tool that introduces a 'formal-driven' paradigm for information retrieval. The framework achieved a score of 60.19 on the GAIA benchmark, surpassing Claude 3.5 Sonnet and GPT-4o. Its innovative data generation method enhances AI's reasoning capabilities in complex tasks, marking a significant advancement in the field.
AMD's Ryzen AI Max+ 395 processor, based on the Zen5 architecture, now supports local operation of AI models with up to 128 billion parameters. This upgrade, achieved through Mixture of Experts (MoE) technology, requires 128GB of unified memory and runs in a Vulkan llama.cpp environment. The processor also handles context lengths up to 256,000 tokens, making high-end AI applications more accessible.
Meta is aggressively pursuing top AI researchers with lucrative offers, including packages exceeding $1 billion, but faces resistance due to concerns about company culture and competition from well-funded startups. Despite personal outreach from Mark Zuckerberg, the tech giant has struggled to attract talent from rivals like Thinking Machines Lab.
Alibaba Cloud has launched 'No Shadow AgentBay,' a cloud platform designed to streamline AI agent development and operation. The platform offers high-performance computing, multi-OS support, and advanced AI capabilities, enabling developers to handle complex tasks with ease. Security and scalability are key features, making it a robust solution for modern AI applications.
The Qwen3-30B-A3B model has released a new version, Qwen3-30B-A3B-Instruct-2507, achieving significant improvements in general capabilities, multilingual support, and long-text understanding. The update positions it competitively against top closed-source models like Gemini2.5-Flash and GPT-4o.
The Shenzhen Institute of Advanced Technology has unveiled the HYPIR Image Restoration Large Model, capable of transforming old photos into 8K ultra-high-definition images in just 1.7 seconds. This breakthrough promises significant advancements in cultural heritage preservation, film restoration, and medical research.
Alibaba's Tongyi Qianwen team launched Qwen3-30B-A3B-Instruct-2507, an open-source AI model that matches GPT-4o and Gemini 2.5-Flash in core capabilities while using only 3B parameters. The model excels in multilingual support, user alignment, and long-text processing, with benchmark scores surpassing competitors in some areas.
OpenAI's GPT-5, featuring multiple versions like GPT-5-pro and GPT-5-high, showcases groundbreaking programming capabilities. Early tests reveal its ability to generate complex websites and enhance classic games, signaling a potential revolution in AI-assisted development. Industry experts highlight its potential to address current AI limitations in complex tasks.
Stepwise Star has launched Step 3, a groundbreaking 321B-parameter multimodal AI model, set to be open-sourced on July 31. The model excels in visual perception and complex reasoning, with significant improvements in efficiency across hardware platforms. The company also announced strategic partnerships to accelerate AI adoption in various industries.
Julius AI, a Y Combinator-backed startup specializing in AI-driven data analysis, has raised $10 million in seed funding led by Bessemer Venture Partners. The platform, which mimics human data scientists through natural language processing, boasts over 2 million users and has generated more than 10 million visualizations. Founder Rahul Sonwalkar highlights the company's focus on practical use cases as key to its success.
Ant Digital has launched Agentar-Fin-R1, a financial reasoning large model built on Alibaba's Tongyi Qianwen Qwen3. The model excels in financial benchmarks and offers tailored versions for diverse institutional needs. It leverages extensive financial data and expert annotations to enhance AI applications in banking, securities, and insurance sectors.
OpenAI's GPT-5, expected in mid-to-late 2025, promises groundbreaking advancements in AI with enhanced reasoning, multimodal capabilities, and reduced hallucinations. This article explores its development, technical specs, and potential applications.
China Mobile has launched its Jiutian Foundation Large Model 3.0, showcasing advancements in language, multimodal, and specialized AI models. The release includes open-source technologies aimed at fostering industry collaboration and innovation.
Zhipu AI has launched GLM-4.5, a groundbreaking open-source model that natively integrates reasoning, coding, and agent capabilities. The model achieves state-of-the-art performance in benchmarks, offers cost-efficient API pricing, and supports full-stack development tasks. Available on Hugging Face and ModelScope, it marks a significant leap in AGI technology.
Chinese AI leader Zhipu AI will launch its next-generation GLM-4.5 model tonight, with details on performance, technical innovations, and open-source plans to be revealed in an online press conference. The global AI community eagerly awaits potential breakthroughs in multimodal capabilities and ecosystem development.
Alibaba's Tongyi Lab has open-sourced WebSailor, an AI agent framework designed to tackle complex information retrieval challenges. The tool has gained rapid popularity on GitHub, outperforming competitors in benchmark tests through its innovative combination of complex task generation and reinforcement learning techniques.
