01
ByteDance Releases Multi-SWE-bench
First Multi-Language Code Auto-Fix Benchmark
The ByteDance Doubao large model team has launched the first multi-language software engineering dataset, Multi-SWE-bench, covering eight mainstream programming languages including Python and Java, with 1,632 real GitHub issue instances. This dataset provides a systematic evaluation of large model code repair capabilities through unified testing standards and professional reviews, promoting the practical development of automated programming and helping developers improve efficiency. (Source: Aibase)
02
Alibaba Cloud Launches MCP Service
Zero-Code Custom AI Agent in 5 Minutes
At the AI Momentum Conference, Alibaba Cloud announced that the Bailian platform has launched the industry’s first full lifecycle MCP service, integrating over 200 large models and more than 50 mainstream MCP services (such as Gaode and Wuying). Users can build a dedicated agent in just 5 minutes without any coding, enabling task decomposition and execution. For example, by combining with Gaode MCP, a travel planning agent can be quickly developed to support weather inquiries and itinerary recommendations. This service has been applied in scenarios such as quality inspection for Kudi Coffee, achieving an accuracy rate of 95%. As of January 2025, over 290,000 enterprise developers are using the Bailian platform, covering 90% of leading industry clients. Alibaba Cloud also announced plans to launch an AI Agent Store to open up ecological agent capabilities. (Source: Quantum Bit)
03
Kimi Open Sources 16B Lightweight Visual Language Model
2.8B Active Parameters Achieve Multi-Modal SOTA
On April 10, 2025, Kimi open-sourced two MoE architecture visual language models, Kimi-VL and Kimi-VL-Thinking, with a total parameter count of 16B but only activating 2.8B during inference. They support 128K long context and exhibit multi-modal reasoning capabilities comparable to models with ten times the parameter count. They perform excellently in tasks such as OCR, geometric problem-solving, and video understanding, surpassing benchmarks like GPT-4o. The core technology includes the MoonViT visual encoder and a three-stage training process (pre-training, supervised fine-tuning, reinforcement learning), particularly enhancing complex reasoning capabilities through long thought chain optimization. The model is now available on Hugging Face, and netizens speculate that its silence period may be paving the way for the upcoming K1.6 model. (Source: Quantum Bit)
04
AI Fitting Rooms Reshape Retail:
Return Rate Decreases by 30%, Online-Offline Integration Becomes a Trend
AI fitting rooms use 3D modeling and dynamic rendering technology to reduce online clothing return rates by 30% and increase live broadcast conversion rates by 50%. Their core value lies in: 1) User body data feedback for design optimization, achieving flexible supply chain management; 2) AR fitting mirrors and other hardware bridging online and offline, creating a full-channel closed loop. After integrating the AI fitting system, Yintai Department Store saw a 26% increase in average transaction value and a 37% decrease in return rates. However, the technology still faces challenges such as privacy security, computational demands, and the lack of industry standards. In the future, it may combine with the metaverse to expand virtual social shopping scenarios. (Source: Hydrogen Consumption)
05
EU Invests €20 Billion to Build AI “Super Factory”
Competing for Global AI Dominance
The EU has announced an investment of €20 billion to build multiple AI “super factories” equipped with over 100,000 processors, focusing on key areas such as healthcare and robotics to close the gap with the US and China in the AI field (with the US and China expected to produce 40 and 15 significant AI models respectively in 2024, while Europe only produces 3). The project will collaborate with private capital to promote green energy and water resource recovery, but environmentalists are concerned about its high energy consumption impacting climate goals. Additionally, the EU plans to develop domestic AI chips and may simplify the AI Act to promote innovation, which has sparked controversy. (Source: AIbase)
06
Google Releases Vertex AI Media Studio:
One-Click Text to Full Video Generation
On April 10, 2025, Google launched Vertex AI Media Studio, integrating four major models: Imagen 3 (image generation), Veo 2 (video dynamicization), Chirp (voice synthesis), and Lyria (background music). Users can automatically generate complete videos by simply inputting text, without needing editing experience. The platform supports shot adjustments and intelligent corrections, with built-in safety filters and digital watermarking technology, balancing efficiency and copyright protection. This tool is aimed at businesses, education, and individual users, and is expected to disrupt traditional video production processes, competing with products like OpenAI Sora in the market. (Source: AIbase)
