AI News

At the 2026 Zhongguancun Forum, Kunlun Tech released its "Tiangong AIGC Family Bucket," featuring three core models: the Matrix-Game 3.0 game world model, the SkyReels V4 video model, and the Mureka V9 music model, all of which are among the world's top-tier. Simultaneously, it announced its "3+1" strategy, using four state-of-the-art (SOTA) models as a foundation to support three AI-native platforms: DramaWave short dramas, Mureka music, and Maosen Academy games, as well as the Skywork super intelligent agent operating system.

SenseTime showcased its AudioClaw input method and SenseAudio AI voice open platform at the Shanghai Global Developer Pioneer Conference from March 27-29. AudioClaw integrates voice input, AI meeting assistant, knowledge base, and other functions, supporting WeChat integration and voice interaction. SenseAudio offers over 70 premium voice options and capabilities such as ASR, TTS, and text-to-image/video processing.

Alibaba DAMO Academy launched its new generation flagship RISC-V CPU IP—XuanTie C950. It achieved a score exceeding 70 in the SPECint2006 benchmark, becoming the world's most powerful RISC-V CPU. The product is the first to achieve native, smooth execution of large-scale models with hundreds of billions of parameters (Qwen3, DeepSeek V3) on a native RISC-V CPU. It integrates a 4K ultra-wide Vector engine and a Matrix engine, achieving a single-core computing power of 8 TFLOPS.

OpenAI has announced the shutdown of its Sora video generation platform, including the app, API, and ChatGPT video functionality. The shutdown is primarily due to OpenAI's strategic downsizing in preparation for its IPO, shifting computing power to next-generation models like Spud and enterprise-level productivity tools. Simultaneously, OpenAI has terminated its three-year IP licensing agreement with Disney and secured a $1 billion investment offer.

Google has launched Lyria 3 Pro, its advanced AI music generation model, which offers significant improvements in generated structure and length compared to its predecessor. The model can accurately handle arrangement elements such as intros, verses, choruses, and bridges, and supports generating complete audio tracks up to approximately 3 minutes long. The model does not directly imitate any particular artist's style; training data uses only legally licensed content, and all generated audio is watermarked with SynthID digital watermarks.

ByteDance's CapCut announced the launch of CapCut Video Studio, integrating the Dreamina Seedance 2.0 audio and video model for its official overseas expansion. The product allows users to create videos from concept to export on an unlimited canvas without a timeline, supporting videos up to 15 seconds long and six aspect ratios. It has already covered Africa, South America, the Middle East, and Southeast Asia.

Suno launches Suno V5.5, introducing a "personalized asset" system for the first time, propelling AI music from a generation tool to a creation platform. Three new features are added: Voices (customizable voices), Custom Models (customizable models), and My Taste (personalized recommendations). Workflow has also been optimized, supporting editing functions such as track separation and partial replacement, resulting in significant improvements in sound quality and emotional expression.

Google has released the Gemini 3.1 Flash Live model, its highest-quality real-time speech model, available in the Gemini App, Search Live, and Google AI Studio. Core upgrades include: speech can directly drive application development (vibe coding), the context window is twice as large, and it supports real-time interaction in over 200 languages across multiple countries. In the ComplexFuncBench audio test, it achieved a function call accuracy of 90.8%, significantly outperforming its predecessor.

Meituan's LongCat team has launched the LongCat-Next multimodal model, with its core innovation being the LoZA sparse attention mechanism. The model supports ultra-long contexts of up to 1MB and intelligently replaces 50% of low-importance modules with streaming sparse attention, achieving a 10x speedup in 128K decoding and a 30% reduction in computational power. The model includes two versions, Flash-Exp and Flash-Lite, and outperforms Qwen-3 in long text tasks.

Alibaba has launched its enterprise-grade agent product, Accio Work, overseas, positioned as "the enterprise version of lobster." Users simply submit their ideas, and the built-in AI agent team can automatically complete the entire process—product selection, procurement, website building, and operation—24/7, generating a viable overseas online store in 30 minutes. The product supports single agent dialogue or agent team group chat modes, allows for customization of roles and skills, and requires user authorization for critical operations to ensure security.

MiniMax has announced the open-sourcing of its production-grade Office document engine, MiniMax Office Skills, comprising four main components: Word, Excel, PDF, and PowerPoint, under the MIT license. This solution addresses the pain point of AI-generated documents being "usable but not user-friendly," ensuring the complete preservation of advanced features such as formulas, formatting, and pivot tables. In terms of technology selection, Word uses the .NET OpenXML SDK instead of python-docx, and Excel directly manipulates XML instead of openpyxl, achieving higher fidelity. (Source: Xiyu Technology)

Alibaba Cloud JVS Claw is now fully open for public beta testing; users can access and use it without an invitation code. The new version adds features such as voice input, 5GB dedicated file space, one-click repair, and Skill toggle, and supports scheduled task management and one-click Clawbot upgrades. New users can enjoy a 7-day free trial, while the paid "Master Edition" is available for as low as 39 yuan for the first month. Users can use the cloud-based OpenClaw service without purchasing Mac mini hardware or incurring API fees.

Alibaba's Tongyi Labs has launched PrismAudio, a video-to-audio generation framework focusing on ambient sound/sound effect synthesis. The model innovatively combines reinforcement learning with a thought chain approach, using a "decompositional thought chain" to allow the model to first analyze video content, timing, sound quality, and spatial location before generating audio. It also incorporates four "teachers" (semantic, temporal, aesthetic, and spatial) to score and optimize the model.

Meta has acquired AI startup Dreamer and its team, founded by Hugo Barra, former global vice president of Xiaomi, and others. Dreamer helps users create AI agents. The Dreamer team will join Meta's Super Intelligence Lab to lead AI agent development. The terms of the agreement have not been disclosed, but investors will receive a return exceeding their investment. Dreamer will retain its independent legal entity status, and Meta will obtain a non-exclusive technology license.

HiDreamClaw, the first native AI agent product for images and videos, has been integrated into Vivago's web platform and made available to overseas users. Based on its self-developed HiDream-I1 multimodal model with over 10 billion parameters, the product incorporates AIGC Skills and supports multimodal creation of images, videos, and text. It also integrates state-of-the-art models such as Nano Banana, Veo, and Sora.

Tencent has launched Ardot, a self-developed AI design tool that helps designers shift from "graphic creation" back to their role as "designers." Ardot aims to be "the most understanding assistant for designers," handling repetitive tasks such as automatically organizing layers and intelligently adjusting details, allowing designers to focus on creativity and insight, and bringing their inspirations to life more quickly.

Meituan has open-sourced the LongCat-Flash-Prover large-scale model, employing a 567.7 billion parameter MoE architecture, focusing on solving complex mathematical proof problems. The model introduces a hybrid expert iterative framework, integrating Lean4 and AST multi-stage rigorous verification to effectively eliminate illusions. It achieved a 97.1% score on the MiniF2F-Test and solved 41.5% of the problems on PutnamBench, both tests setting new state-of-the-art (SOTA) records.

Zhipu has launched GLM-Claw intelligent agents, enabling the creation of a team of "AI assistants who understand you" to help handle various tasks. Positioned as a personalized AI assistant team, the product can collaboratively complete complex tasks, improving users' work and life efficiency. Through a multi-agent collaboration mechanism, it upgrades from single-conversation to multi-task processing.

WeChat has officially launched the WeChat ClawBot plugin, supporting integration with OpenClaw (Lobster). Users can obtain installation instructions through "Me - Settings - Plugins," and scan the QR code to connect Lobster to WeChat for chat-based remote control. The plugin is currently being rolled out gradually and requires an update to the latest WeChat version. Meanwhile, Tencent Cloud Lighthouse (including the enterprise version Claw Pro), the self-developed WorkBuddy, and local QClaw also support WeChat channels.

Baidu AI Cloud DuMate (Chinese name: 搭子) has officially launched, becoming the first enterprise-grade, full-fledged version of OpenClaw in China. The product uses natural language as its interaction interface, supports native operations of office software such as Word, Excel, and PowerPoint, and boasts three core functions: intelligent file management, multi-source data analysis, and office operation automation. DuMate integrates the entire suite of Baidu Skills capabilities, including Baidu Search, and supports flexible expansion.