AI News

NVIDIA released the Nemotron3Embed embedding model series for production-grade RAG, agent retrieval, code search, and memory. The 8B model ranks first on the RTEB benchmark, making it the top open-source embedding model. The series includes three checkpoints: accuracy-focused 8B-BF16, lightweight 1B-BF16, and 1B-NVFP4 4-bit optimized for Blackwell architecture. All models use bidirectional attention.....

At the 2026 World AI Conference, Dewu App and Qianjue Robot launched the world's first AI authentication robot, exploring embodied intelligence for product authentication. Leveraging billions of inspection data from its 'authenticate first, ship later' model, Dewu developed an 'AI brain' authentication system achieving precise verification through fine-grained perception.....

On July 17, Chongqing Health Commission launched "Yuxiaojian", an AI-assisted medical agent for residents. Accessible via Alipay or Afu App, it offers smart triage, cloud escort, report query & AI interpretation. Jointly built with Big Data Bureau & Ant Group, powered by digital Chongqing infrastructure, hospital data and Ant AI model.....

Galaxy General Robot CTO Wang He predicted at the 2026 World AI Conference that embodied intelligence will achieve a major breakthrough before 2028, with performance comparable to ChatGPT. The foundational model, trained on massive data, can reach a 70%-80% success rate on tasks not specifically trained for, similar to early digital models.....

Wan-Streamer v0.2 from Tongyi Lab eliminates lag, delay, and audio-video desync in video calls. This end-to-end multimodal model responds in just 550ms, integrating listening, seeing, speaking, and acting within a single Transformer for near-real face-to-face natural interaction.....

Top tech vlogger Film Storm launched a 49-yuan AI course, selling 100k+ copies on day one with 4.9M yuan revenue. Originally 99 yuan, the 16 lessons cover AI image/video generation, prompts, and practical film effects like robotic arm moves, bullet time, and AI storyboarding.....

WeRide launches WeRide WITT, a physical AI cognitive foundation model. Its core introduces the 'minimum physical fact unit' concept, enabling AI to better understand multimodal data like video, images, and text, improving cognition for complex autonomous driving scenarios.....

WeRide unveiled its WIIT large model at WAIC 2026, introducing the 'minimal physical fact unit' concept. It breaks continuous environments into foundational facts, building a framework with fact extraction, reasoning, and verification modules, advancing AI from data understanding to real-world cognition.....

StepFun at WAIC2026 showcased its first large-model native agent phone, STEPX Neo, with orange design. The system includes modules like "Seafood Market" and can integrate third-party agents via partnerships, exploring deep AI agent integration with mobile terminals.....

Baidu's Shen Dou predicts a surge in general and industry-specific AI agent deployments by H2 2026. The next three years are key, with 90% of work deeply augmented by AI, not fully replaced.....

On July 16, Kunlun Wanwei launched the Tiangong Short Drama Workbench, noting that while AI short dramas are moving toward high quality, blind-box generation still faces three major issues: uncontrollable creation (character face shifts, position drift, narrative fractures); difficulty balancing quality and output (fully automated generation feels robotic, manual refinement can't scale).....

The main battlefield of large model services is shifting toward retrieval-augmented question answering, multi-document summarization, and long-range agents. The request prompt is composed of dozens to hundreds of semantically independent segments (retrieved documents, skills explanations, memory, historical rounds), forming ultra-long context with tens of thousands to hundreds of thousands of tokens. The pre-filling stage dominates the computing costs, becoming the most prominent cost source for service providers, and triggering more challenging problems.

At WAIC 2026, Soul launched B Soul, an AI hardware showcasing real-time multimodal interaction and emotion perception. CTO Tao Ming said the company evolved from a social app into an ecosystem focused on emotion sensing, interaction tech, and self-developed large interaction models, distinct from general-purpose LLMs.....

OpenAI confirms that Codex powered by GPT-5.6, with Full Access Mode and no sandbox, may delete home directory files by overwriting $HOME variable. Lead Tibo Sottiaux reports few cases, triggered only under specific conditions.....

On July 16 at WAIC Shanghai, Baidu AI Cloud's no-code platform Miaoda released version 3.5, evolving from 'one-sentence app creation' to 'simpler and more comprehensive,' further lowering development barriers. It has served 35 million users, created 3.5 million commercial apps, with nearly 200,000 daily active users.....

Zhipu AI spent hundreds of millions of yuan acquiring AI infra firm Zhongke Jiahe to strengthen domestic computing adaptation and inference optimization. Originating from CAS's ICT, it specializes in compilation tech, with experience in compilers for Loongson and Huawei Ascend. It provides full-chain capabilities from virtual instruction sets to deployment, filling model low-level engineering gaps.....

Mingshi AI opensources the enterprise-level platform StaffDeck. Unlike chatbots, it focuses on digital workforce. It can transform organizational knowledge, SOPs, and decision rules into digital employees. The platform, through code open source (GitHub repository), aims to build and manage digital workforces capable of performing professional tasks.

Roblox launched AI creation tool Build and upgraded its Studio, using AI to lower the development barrier. Users can generate editable game content via text prompts. The feature begins testing on July 28, deepening the 'user-generated content' philosophy. The platform has 132 million daily active users. Build is a mobile-first tool, enabling creation anytime, anywhere.....

JD.com launched the Jingmai AI Business Center, an AI-driven platform integrating AI tools, expert services, and intelligent management to upgrade merchants' operations. It opens 22 core AI tools, leveraging multi-agent collaboration, A2A smart interaction, and natural language processing to empower all-link e-commerce intelligence.....

Traditionally, switching programming languages takes years, a major disruption for companies. Anthropic used Claude (Fable5, Opus4.8) and dynamic workflows to smoothly migrate 10 codebases (tens of thousands to hundreds of thousands of lines each) to a new language in just one month, overturning conventional wisdom and demonstrating the efficiency and viability of AI-driven large-scale code refactoring.....