Welcome to the 【AI Daily】 column! Here is your guide to exploring the world of artificial intelligence every day. We present you with the hottest topics in the AI field, focusing on developers to help you gain insights into technology trends and innovative AI product applications.
New AI products click to learn more: https://top.aibase.com/
1. DeepSeek Launches FlashMLA: A Powerful Tool for Large Model Acceleration with Decoding Performance Boosted to 3000GB/s
On the first day of Open Source Week, DeepSeek introduced FlashMLA, an efficient multi-layer attention decoding kernel designed for NVIDIA's Hopper architecture GPUs, aimed at optimizing large model inference performance in variable-length sequence scenarios. This technology significantly enhances memory management accuracy by fully supporting BF16 precision and an innovative page-based key-value caching system. FlashMLA demonstrated excellent performance in tests, achieving a processing speed of 3000GB/s and a computing power of 580TFLOPS.
【AiBase Summary:】
⚙️ FlashMLA is designed for NVIDIA's Hopper architecture GPUs, optimizing variable-length sequence scenarios and enhancing inference performance.
💾 Utilizes BF16 precision and a page-based key-value caching system with a block size of 64 for precise memory management.
📈 On the H800SXM5 GPU, FlashMLA achieves a processing speed of 3000GB/s in memory-constrained scenarios.
Details link: https://github.com/deepseek-ai/FlashMLA
2. Hailuo AI Unveils New "Director" Mode: I2V-01-Director Model Now Available to Everyone
The latest I2V-01-Director model from Hailuo AI offers a groundbreaking AI video generation tool for global users, marking another significant advancement in AI for film creation. This tool allows users to control camera movements through natural language commands, easily transforming ideas into cinematic works.
【AiBase Summary:】
📽️ The I2V-01-Director model supports natural language commands, enabling users to control camera movements with simple descriptions.
🎨 This tool lowers the barrier for professional video production and enhances the creative freedom of creators.
🌐 Hailuo AI has integrated this model into its video creation platform for immediate user experience.
Details link: https://hailuoai.video/create
3. Google's AI Video Generation Model Veo 2 Pricing Revealed: $30 for One Minute of Video
Google recently announced the pricing information for its video generation AI model Veo 2, attracting widespread attention. The usage fee for Veo 2 is $0.50 per second, which translates to a cost of $30 for one minute of video and up to $1800 per hour. Compared to the $32,000 per second production cost of Marvel's "Avengers: Endgame," Veo 2's pricing appears relatively affordable. 
【AiBase Summary:】
💰 The video generation cost for Veo 2 is $0.50 per second, with a maximum hourly cost of $1800.
🎬 Compared to the $32,000 per second production cost of Marvel's "Avengers: Endgame," Veo 2's pricing is quite competitive.
💡 OpenAI's Sora model subscription costs $200 per month, giving Veo 2 a competitive edge in the market.
4. Grok 3 Launches Real-Time Voice Feature, Supporting 10 Interaction Modes
The Grok APP recently introduced a real-time voice mode, supporting up to 10 interaction modes, enhancing user experience. The newly added Unhinged and Sexy modes have garnered significant attention, with the former characterized by provocative debates and the latter filled with romantic implications. This update not only enriches conversation methods but also raises discussions about content moderation and its impact on underage users.
【AiBase Summary:】
🎙️ The Grok APP introduces a real-time voice mode, supporting 10 interaction styles to enhance user experience.
🔥 The new Unhinged and Sexy modes offer provocative debates and flirty interactions.
🔍 This update has sparked discussions on content moderation and the effects on underage users, as xAI needs to balance innovation and safety.
5. Baidu APP Gets a Complete Revamp: Launches AI Entry DeepseekR1 for Seamless Search
Baidu App today announced the full launch of a new AI entry, deeply integrating the Wenxin large model and DeepSeek model to provide users with a fully upgraded experience. New features include AI search, comprehensive web information extraction, AI summarization, and digital human agents, which can accurately understand user needs and deliver personalized search results. Leveraging vast user data and large model inference capabilities, Baidu App presents multimodal content recommendations to enhance information retrieval efficiency.
【AiBase Summary:】
🔍 New features such as AI search and comprehensive web information extraction enhance user experience.
👤 By analyzing user profiles and behavior data, personalized search results are provided, achieving a truly tailored experience.
📈 On the first day of the upgrade, usage exceeded ten million times, indicating high user recognition of the new features.
6. Tencent Hunyuan Video Model Testing Recruitment Underway, Open Source on the Horizon
Tencent's Hunyuan Video model has recently opened recruitment for testing partners on X platform, marking a critical testing phase for its AI video generation technology. With over 13 billion parameters, this model stands out among open-source video generation models, showcasing exceptional performance, particularly in video smoothness and multi-object generation.
【AiBase Summary:】
🎯 The Hunyuan video model is a significant innovation from Tencent in the AI video generation field, with over 13 billion parameters and powerful generation capabilities.
🛠️ This model has shown excellent performance during testing, particularly excelling in video smoothness, detail restoration, and multi-object generation compared to existing models.
🤝 The recruitment of testing partners aims to enhance model performance through collaboration, with participants providing feedback on stability and usability before the model is open-sourced.
Details link: https://docs.google.com/forms/d/e/1FAIpQLScZe9DflcpKf_JfcgGnAiXfyhEZAt2uP4mE0Rhs5l9R2m49oA/viewform
7. Moonshot AI Collaborates with UCLA to Launch New Model Mixture-of-Expert, Enhancing Language Model Training Efficiency
In the field of artificial intelligence, the Moonlight model developed jointly by Moonshot AI and UCLA employs the Muon optimizer, significantly improving the training efficiency and stability of large language models. Through innovative Newton-Schulz iteration and weight decay techniques, Moonlight outperforms traditional AdamW optimizers across multiple tasks, particularly excelling in language understanding and code generation.
【AiBase Summary:】
🌟 The Moonlight model, a Mixture-of-Expert model developed by Moonshot AI and UCLA, offers configurations of 3 billion and 16 billion parameters, trained on 5.7 trillion tokens.
⚙️ The Muon optimizer significantly enhances training efficiency and stability for large models through Newton-Schulz iteration and weight decay techniques.
📈 Empirical results show that Moonlight outperforms traditional AdamW training models across various tasks, demonstrating superior language understanding and code generation capabilities.
Details link: https://github.com/MoonshotAI/Moonlight?tab=readme-ov-file
8. 1X Launches Home Robot NEO Gamma: Capable of Making Coffee, Doing Laundry, and Vacuuming
The latest home robot Neo Gamma from 1X Company comes equipped with various household functions such as making coffee, doing laundry, and vacuuming, while also providing companionship to users. Unlike traditional home robots, Neo Gamma is designed to be more gentle and approachable, easily fostering emotional attachment. Its advanced motion control system and language interaction capabilities enable it to seamlessly integrate into family life, becoming a reliable assistant.
【AiBase Summary:】
🌟 Neo Gamma is a new type of home robot capable of completing household chores while providing companionship.
🤖 Its design is more friendly and gentle, unlike traditional home robots, making it easier to develop emotional connections.
🛠️ Neo Gamma possesses natural movement abilities and language interaction features, which will closely integrate with family life in the future.
Details link: https://www.1x.tech/neo
9. Xiaomi to Launch First AI PC Product
Xiaomi Group partner and president Lu Weibing announced during a recent live stream that Xiaomi is about to launch its first AI PC product, drawing widespread attention in the tech community. The standout feature of this new product is its 99Wh ultra-large battery, which not only meets civil aviation transportation safety standards but also significantly enhances user endurance during mobile office use. With the increasing popularity of mobile office work, user demand for battery life is rising, and Xiaomi's design directly addresses this issue.
【AiBase Summary:】
🔋 The new AI PC features a 99Wh ultra-large battery, meeting civil aviation transportation safety standards and making it a carry-on product.
💼 This product aims to enhance user endurance during mobile office work, alleviating battery anxiety.
🤖 Deep integration of AI technology provides a smarter operating experience, enhancing convenience for work and study.
10. Alibaba Announces Record Investment: 380 Billion RMB for Cloud and AI Hardware Development Over the Next Three Years
Alibaba Group CEO Wu Yongming announced that the company will invest over 380 billion RMB in cloud and AI hardware infrastructure over the next three years. This record investment is not only the largest by a private enterprise in China in this field but also demonstrates Alibaba's technological ambition and strategic layout in the AI era. 
【AiBase Summary:】
🚀 Alibaba will invest 380 billion RMB in cloud and AI hardware infrastructure over the next three years.
📈 The domestic AI industry is experiencing explosive growth, and Alibaba's investment will accelerate the industry's intelligent transformation.
🌍 This investment will enhance China's competitiveness in the global AI race, helping enterprises secure important positions on the international stage.
11. Meta AI Releases New Video Learning Model V-JEPA: A Breakthrough in Video Understanding
The Meta AI team has recently launched the V-JEPA model, achieving a significant breakthrough in unsupervised video learning. This model learns directly from video data using a feature prediction principle, avoiding the reliance on pre-trained encoders and text supervision found in traditional methods. V-JEPA excels in motion understanding and video tasks, demonstrating its efficient training capabilities and strong representation abilities, particularly maintaining accuracy even in low-sample settings.
【AiBase Summary:】
🔍 The V-JEPA model focuses on unsupervised feature prediction, learning directly from video data and avoiding the limitations of traditional methods.
⚡ This model performs excellently in video tasks and low-sample learning, showcasing efficient training capabilities and strong representation abilities.
📈 V-JEPA has an advantage in capturing subtle motion details, proving its potential in video understanding.
Details link: https://ai.meta.com/research/publications/revisiting-feature-prediction-for-learning-visual-representations-from-video/
12. LiblibAI Secures Hundreds of Millions in Funding: Completing Four Rounds of Financing Within a Year
LiblibAI has successfully completed four rounds of financing within just one year, setting a new record for the speed of funding in the domestic AI application sector. This latest round secured hundreds of millions in funding, further consolidating its leading position in the industry. The company is dedicated to transforming the creative process of creators through AI technology, establishing a complete creative ecosystem, and continuously focusing on user growth and technological research and development.
【AiBase Summary:】
💰 LiblibAI has completed four rounds of financing in one year, setting a new record for the speed of funding in the domestic AI application sector.
🖼️ The company's AI model/image creators have surpassed 20 million, with daily interactions reaching millions.
🌐 This round of financing will focus on building the creator ecosystem and technological research and development to enhance platform technical strength and user experience.