LiblibAI recently completed a $130 million Series B funding round, co-led by Sequoia China, CMC Capital, and major industry investors, marking the largest funding round in China's AI application sector this year. Last month, LiblibAI released version 2.0, evolving from a model community into an AI creation studio.
Shenzhen Interstellar Lightyear Technology Co., Ltd. launched its new product, the Pantheon 22, a five-finger dexterity hand, and completed its Pre-A round of financing. This round was led by Seine Capital and Puhua Capital, with participation from Kexi Venture Capital and DeepMind Capital serving as financial advisor. The funds will primarily be used for tackling key technical challenges related to dexterity operation, developing the dexterity hand operating system (cerebellar model), and expanding the core team.
Claude desktop version has been officially released, supporting Mac and Windows systems. The version has four core functions: global keyboard shortcuts (double-click Option on Mac to wake up), work sharing (screenshot, window sharing, drag and drop files), voice input (press Caps Lock to speak), and connection tools (can access code editor, local files, and databases).
The Qwen3-VL family adds two new dense model sizes, 2B and 32B, ranging from lightweight to mid-range, covering visual language understanding scenarios. Two versions are available: Instruct is suitable for dialogue and tool calls, while Thinking enhances long-chain inference and complex visual understanding. The Qwen3-VL-32B outperforms models like the GPT-5 mini in multiple domains, matching the performance of a 235B model with only 32B parameters. The Qwen3-VL-2B, despite its small size, delivers impressive performance and can run on extreme edge devices.
Beijing-based KuaFuAI has officially open-sourced AipexBase, China's first AI-native backend-as-a-Service platform. This platform allows developers to "have a complete backend without writing their own," as all backend capabilities, such as data storage and authentication, are automatically encapsulated and can be invoked with a single click via the frontend SDK or the MCP protocol. AipexBase is natively compatible with MCP, unifying the context and data layer, and supports integration with AI products such as CodeFlyer and Cursor. It is deeply adapted to the Chinese development ecosystem and compatible with ecosystems like Lark, DingTalk, and WeChat.
Tencent has released and open-sourced the Hunyuan World-Mirror model 1.1. The model now supports multiple views and video input, can be deployed on a single GPU, and generates 3D worlds in seconds. Breaking through the limitations of version 1.0, which only supported text or single-image input, it is the first to achieve end-to-end 3D reconstruction with multimodal prior injection and unified output across multiple tasks. It supports various 3D geometric predictions, including point clouds, depth maps, and camera parameters, and its performance significantly outperforms existing methods.
OpenAI has released its first AI-native browser, ChatGPT Atlas. Built on Google's open-source Chromium engine, the browser integrates ChatGPT conversational capabilities, allowing direct chat across all tabs. It features a browser context assistant, enabling users to ask questions directly on the page; a built-in memory function to record key browsing content; a "Cursor Chat" feature that allows editing and polishing of Chinese text; and an Agent mode for completing web page operations such as filling out forms and making reservations.
Meituan's LongCat team released VitaBench, a benchmark for evaluating large-scale intelligent agents facing complex problems. Using three major life scenarios—food delivery, restaurant dining, and travel—as platforms, it constructs an interactive evaluation environment containing 66 tools. VitaBench quantifies task complexity from three dimensions: deep reasoning, tool usage, and user interaction, achieving fine-grained evaluation through a real user simulator and atomic evaluation criteria.
Zhipu has released the GLM Coding Plan Enterprise Edition, based on the GLM-4.6 model, providing enterprises with a comprehensive intelligent programming solution. The product ranks first in the OpenRouter trending list on the international API platform, integrating multimodal understanding, network search, and intelligent orchestration capabilities to provide a one-stop service from code generation to end-to-end development collaboration. The Enterprise Edition features high usage, low cost, high performance, and high security, seamlessly adapting to more than 10 mainstream programming tools and supporting flexible member management and data analysis.
Cherry Studio Enterprise Express Edition has been officially released, offering a one-stop AI implementation solution for enterprises at a buy-out price of 5000 RMB with no limit on employee seats. This version includes top-level closed-source and open-source models, eliminates the need for API Key applications, resolves network issues, simplifies billing processes, and supports private and cloud application deployments. Studio - Qianhui Technology
CodeBuddy IDE 1.0 official version released. The international version fully supports GPT-5-Codex, and adds features such as custom commands and Web Fetch to obtain network information in real time. The MCP Marketplace is officially open and supports full DiffView functionality.
Anthropic has released a web version of Claude Code, allowing users to utilize its code generation capabilities directly in their browsers without requiring deployment. Core features include connecting to GitHub repositories, automated code writing, parallel processing of multiple tasks, real-time progress tracking, and automatic creation of pull requests (PRs) upon completion.
Vidu Q2 Reference Videos are now available globally. This upgrade focuses on three core aspects: high consistency, faster speed, and more affordable pricing, meeting the growing demand for highly imaginative content creation from professional and semi-professional creators. Vidu Q2 Reference Videos are currently available worldwide. Users can experience the latest features on the Vidu website or by searching for the Vidu AI APP in major app stores.
Baidu's self-developed multimodal document parsing model, PaddleOCR-VL, topped the HuggingFace Trending charts globally within just 16 hours of its release. With only 0.9B parameters, the model is lightweight and efficient, accurately recognizing various complex elements such as text, handwritten Chinese characters, and tables, and supports 109 languages.
Unitree Robotics has released its fourth humanoid robot, the H2, which stands 180cm tall and weighs 70kg. Compared to its predecessor, the H1, the H2 shows significant improvements in motion fluidity and biomimetic features. The H2 has 31 joints distributed across its arms, legs, and torso, enabling it to perform complex movements such as ballet and Chinese martial arts. It also features human-like facial features, making its appearance and movements more human-like.
DeepSeek introduces its new visual text compression model, DeepSeek-OCR. With only 3 bytes of parameters, a hybrid expert architecture, and a 20x reduction in visual tokens, it achieves a 20x compression ratio and can process 33 million pages of data per day across 20 nodes. In Fox benchmark tests, it achieves over 85% accuracy across various text length ranges. It supports multiple resolution configurations, multilingual processing, complex chart parsing, and other multimodal capabilities, achieving 10x compression efficiency in multi-turn dialogues.
AI video company PixVerse announced the completion of a RMB 100 million Series B+ financing round, jointly invested by Fosun RZ Capital, Tongchuang Venture Capital, and Shunxi Capital. PixVerse and PixWo AI, PixVerse's products, serve both the general public and professional creators, boasting a user base exceeding 100 million, an ARR of over $40 million, and a monthly active user base of over 16 million.
Manus has announced the release of Manus 1.5. It features significant improvements in task execution speed, reliability, and output quality, reducing the average task completion time to less than 4 minutes, a nearly fourfold increase in speed. Manus 1.5 offers two models: Manus-1.5 for highly complex tasks and Manus-1.5-Lite for cost-efficiency optimization. The new version also features full-stack web application development capabilities, allowing users to complete the entire process from development to deployment through dialogue.
Meituan's LongCat team has open-sourced its LongCat-Audio-Codec speech codec solution. Designed specifically for large speech language models (Speech LLM), it employs a parallel extraction mechanism of semantic and acoustic dual tokens to balance the semantic and acoustic features of speech, solving the problem of balancing semantic and acoustic information in traditional solutions. The low-latency streaming decoder supports real-time interaction, meeting the needs of scenarios such as in-vehicle voice assistants and real-time translation.
World Labs, a startup founded by Stanford University professor Fei-Fei Li, has launched a real-time generative world model, RTFM. The model can run on a single H100 GPU, and after taking a 2D image as input, it can generate new 2D images from different perspectives, achieving 3D consistency and persistence. RTFM is trained on large-scale video data and does not require explicit construction of 3D representations; instead, it learns and simulates 3D geometry, reflection, and other features through end-to-end learning.
Logic Robotics has released its new generation of industrial-grade interactive embodied robot – the Logic Elf G2. Built to industrial standards, the robot is equipped with an NVIDIA Jetson Thor chip, high-precision force-controlled dual arms, and 19-DOF dexterous hands. It features 3D tactile perception and a 5-DOF waist and legs combined with an omnidirectional chassis. It supports multi-person continuous voice dialogue and knowledge base question answering, employs dual-battery hot-swappable battery swapping technology, and is equipped with a 360° surround-view fisheye camera and front and rear dual LiDAR sensors for active obstacle avoidance.
Alibaba has launched Qoder CLI, a new AI programming tool designed specifically for command-line environments. Integrating top-tier programming models and a lightweight agent framework, it boasts powerful code generation and understanding capabilities while reducing memory consumption and command response time, thus improving development efficiency. Qoder CLI requires no complex initialization, is ready to use upon installation, supports file editing, command execution, and other functions, and can be extended or customized via MCP.
Anthropic released the Claude Haiku 4.5 model. While maintaining high performance, the model doubles in speed and significantly reduces cost. In the SWE-bench Verified test set, Haiku 4.5 achieved a score of 73%, placing it on par with Claude Sonnet 4 and OpenAI's GPT-5, and even surpassing Sonnet 4 on some tasks.
Volcano Engine has released and upgraded four new Doubao large models, including Doubao Large Model 1.6 Upgrade, Doubao Large Model 1.6 Lite, Doubao Speech Synthesis Model 2.0, and Doubao Voice Replication Model 2.0. Doubao Large Model 1.6 Upgrade natively supports four thought lengths, making it the first model in China to natively support "tiered adjustment of thought length," balancing performance, latency, and cost. Doubao Large Model 1.6 Lite is lighter, has faster inference speed, and offers better cost-effectiveness.