Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we bring you hot topics in the AI field, focus on developers, and help you understand technology trends and innovative AI product applications.
Fresh AI products Click to learn more: https://app.aibase.com/zh
1. Moonshot AI Makes a Big Move! Kimi Hosted Agents Officially Released
Moonshot AI has moved the agent runtime framework to the cloud, providing an out-of-the-box, around-the-clock hosted environment that significantly lowers the operational threshold for long-task execution and multi-agent collaboration. This is a pragmatic exploration of productizing AI Agent infrastructure.
[AiBase Summary:]
☁️ Harness goes cloud-based, eliminating the need for enterprises to build and maintain their own runtime framework
🕒 Provides a 7×24 hosted environment to support stable execution of long, complex tasks
🤝 Provides a stable foundation for multi-agent collaboration and lowers the barrier to deploying advanced AI applications
2. Google Makes a Major Release with Omni 1.2 Flash: Video Generation Extended to 15 Seconds, API Price Cut by 30%
Google DeepMind's release of Omni1.2 Flash forms a combined upgrade around duration, consistency, multi-turn editing, and pricing. The 15-second duration and draft-upscaling strategy balance efficiency and cost, while simultaneous launches across platforms show ecosystem coordination. However, actual generation quality and consistency across third-party platforms still need ongoing verification.
[AiBase Summary:]
🎬 Generation duration increased to 15 seconds, supporting custom lengths from 3 to 15 seconds
🧠 Multi-turn conversational editing can precisely replace subjects, backgrounds, and lighting
💰 720p API price reduced by about 30%, with a 15-second cost of about $1.05
3. Taobao MNN Team Open-Sources MNN-Pocket-0.3B, Offline Decoding on iPhone 18 Pro Soars to 350 tokens/s
With 0.3B parameters, this model balances the most critical factors for on-device deployment: size, memory, speed, and power consumption. INT4 deployment is only 173MB, peak memory increase is 228MB, and it reaches up to 466 tokens/s on Apple devices, giving it practical value. The engineering optimization approach is clear, but the parameter scale means it is better suited for lightweight tasks and cloud-device collaboration.
[AiBase Summary:]
🧠 Lightweight design: INT4 size 173MB, peak memory increase 228MB, low power consumption
⚡ Performance beyond its class: about 350 tokens/s offline on iPhone, surpassing peers in 11 evaluations
📦 Full-process open source: multiple versions covering Base/SFT/MNN on-device deployment
4. Generative AI Image Editing Accelerates Its Push Downmarket as WhatsApp Tests Smart Outpainting on Mobile
Generative AI is rapidly penetrating from professional tools into instant messaging scenarios. WhatsApp's move validates the feasibility of lightweight, scenario-based AI image editing. The feature and interaction design are simple, lowering the user threshold, but generation quality and privacy boundaries still need to be observed.
[AiBase Summary:]
📱 WhatsApp iOS beta launches AI outpainting, integrated into the drawing editor
🖼️ Users can drag beyond the frame to automatically generate extensions, with support for undo and regeneration
🚀 Professional-grade image editing features are moving down into social scenarios, lowering the barrier to entry
5. Samsung Invests 4.2 Billion to Build a Data Center, Keeping Computing Power for Its Own People First
Samsung SDS has begun construction of an AI data center in Gumi, with an initial 60MW and later 120MW, representing an investment of about RMB 2.131 billion. Its core purpose is not simply taking external orders, but providing stable computing power for the group's manufacturing robotics transformation. Building one's own computing power is becoming standard for conglomerates seeking to control the pace of transformation.
[AiBase Summary:]
🏗️ Construction of an AI data center begins in Gumi, initially 60MW and expandable to 120MW
🤖 The core use is to provide computing power support for the group's manufacturing robotics transformation
🔒 Self-built computing power is becoming standard for conglomerates, ensuring the pace of transformation is not controlled by others
6. Perplexity Launches a New "Teammates" Feature: Managing Ongoing Projects with AI Agents
Perplexity is pushing AI agents from single tasks to ongoing work management. Through a multi-model foundation and project memory, it enables cross-tool coordination, while isolated environments enhance enterprise-grade security and permission control. This feature marks the transformation of AI assistants from passive tools into operational agents capable of autonomous planning and execution, with a clear path to deployment.
[AiBase Summary:]
🚀 Launches Teammates, aimed at ongoing complex work management
🔗 Cross-tool coordination and project memory to handle follow-ups
🔒 Isolated environments and access control, available with a Max subscription
7. Giving AI a Face and a Voice: Meta's Muse Can Now Video Chat
From a product implementation perspective, Meta combines speech synthesis and digital human driving into a real-time model pipeline. An initial response time of about 870 milliseconds and an interruptible design improve conversational naturalness. However, being limited to U.S. users, computing costs, and multilingual adaptation remain deployment challenges.
[AiBase Summary:]
📞 Muse launches real-time voice and virtual avatars, supporting interruptions at any time
⚡ Dual models generate speech, expressions, and lip sync in real time, with about 870 milliseconds of latency
📱 Aimed at scenarios such as walking or being busy, allowing brainstorming and sharing to social media
8. 47,000 Fake Celebrity Content Removed: Cyberspace Administration Hits the Brakes on AI Face-Swapping for Sales
Regulation is shifting from case-by-case cleanup to normalized special governance, precisely targeting monetization gray industries such as AI face-swapping for sales and fake personas for traffic, while tightening platform responsibility. However, cross-platform coordination and traceability mechanisms for AI-generated content still need strengthening to prevent the black market from reviving.
[AiBase Summary:]
🧹 Cleaned up 47,000 pieces of fake celebrity AI information and handled 3,961 accounts
💰 All four typical categories aim at monetization: stealing portraits for sales, spreading rumors, impersonating viewpoints, and attracting followers with vulgar content
⚖️ Special actions continue, tightening platform responsibility and narrowing outlets for AI abuse
