Welcome to the "AI Daily" column! This is your guide to exploring the world of artificial intelligence every day. Every day, we present you with the latest content in the AI field, focusing on developers to help you understand technical trends and learn about innovative AI product applications.
Fresh AI products click to learn more:https://app.aibase.com/zh
1. Tencent's major release of the Hunyuan 3D 3.0 model, improving modeling accuracy by 3 times
Tencent officially released the Hunyuan 3D 3.0 model at the 2025 Global Digital Ecosystem Conference. Its 3D-DiT hierarchical carving technology significantly improves modeling accuracy, while also launching the Hunyuan 3D Studio platform and open-source plan, pushing the boundaries of 3D creation.
AiBase Summary:
🧠 The Hunyuan 3D 3.0 model uses 3D-DiT hierarchical carving technology, improving modeling accuracy by 3 times.
🎨 The Hunyuan 3D Studio platform provides professional creative tools, enhancing the efficiency and quality of 3D art creation.
🚀 Tencent plans to open-source the Hunyuan 3D omni model, accelerating the application of 3D generation technology in academic and industrial fields.
2. Kunlun Wanshi AI Music Creation Platform Mureka Launches Agent Studio Function, Making Music Creation Accessible!
Mureka's "Agent Studio" function makes music creation accessible through an intuitive approach. Users just need to describe their ideas simply, and the AI will automatically generate lyrics and music. This feature includes multiple creative scenarios, such as album production and hot song writing, providing users with diverse music experiences.
AiBase Summary:
🎧 Mureka introduces the "Agent Studio" function, making it easy for ordinary people to create music.
🤖 Users just need to state simple ideas, and the AI can generate complete lyrics and music.
🎶 There are currently six creative scenarios, covering album production, hot song writing, emotional expression, and more.
3. Alibaba Qoder Launches Paid Subscription Service, Pro Version Costs $20 per Month, Helping AI Self-Programming
Qoder officially launched a paid subscription plan, offering Pro and Pro+ versions, supporting unlimited code completion, advanced model calls, and other features to improve development efficiency. At the same time, it optimized the credit consumption issue, improving the parallelization ability of the intelligent agent tools and the accuracy of engineering retrieval.
AiBase Summary:
🔥 Qoder launches a paid subscription service, supporting Pro and Pro+ versions to meet developers' needs for efficient programming.
💡 Pro version provides unlimited code completion and 2000 Credits, while the Pro+ version offers 6000 Credits and more resources.
🚀 Optimized credit consumption, improved the parallelization capability of intelligent agent tools, and reduced token consumption.
4. VEED Fabric 1.0 Released! One Picture Becomes a "Talking" Video
VEED's Fabric 1.0 is a revolutionary AI video generation tool that can produce high-quality talking videos from a single image and voice input. This tool excels in lip synchronization, natural facial expressions, and generation speed, significantly reducing the cost and time of video production, suitable for various content creation scenarios.
AiBase Summary:
🖼️ Fabric 1.0 supports generating lively talking videos from static images, achieving dynamic storytelling.
⏱️ Video generation speed increased by 7 times, cost reduced by 60 times, suitable for fast content production.
🌐 Integrated multilingual support and automatic subtitle functions, enhancing the user experience for global users.
Details link: https://www.veed.io/ai/fabric-1-0
5. OpenAI Launches GPT-5-Codex: AI Coding Agent Will Completely Revolutionize the Developer World
OpenAI released GPT-5-Codex, marking a significant breakthrough in the AI agent coding field, with its dynamic thinking mechanism and multi-platform integration capabilities significantly improving software development efficiency.
AiBase Summary:
🧠 GPT-5-Codex has a dynamic thinking mechanism, which can adjust processing time according to task complexity, improving coding efficiency.
💻 Supports multi-platform integration, including IDE extensions, web interfaces, and GitHub code review features, enhancing the developer ecosystem.
🚀 Developer feedback shows that GPT-5-Codex significantly shortens the development cycle, improves code generation speed, and reduces error comments.
Details link: https://openai.com/index/introducing-upgrades-to-codex/
6. National Release of "Artificial Intelligence Security Governance Framework" 2.0 Edition, Promoting the Construction of a Safe and Trustworthy AI Ecosystem
The "Artificial Intelligence Security Governance Framework" 2.0 edition was officially released on September 15, 2025, aiming to address new challenges brought by the rapid development of AI technology. The framework optimizes the 1.0 edition based on practical applications, improves risk classification and prevention measures, and emphasizes the importance of global cooperation.
AiBase Summary:
🔐 The "Artificial Intelligence Security Governance Framework" 2.0 edition was officially released to address new challenges brought by AI technology development.
🔍 The framework optimizes the 1.0 edition, improving risk classification and prevention measures.
🤝 Emphasizes global cooperation, promoting AI security governance cooperation under multilateral mechanisms.
Details link: https://www.cac.gov.cn/2025-09/15/c_1759653448369123.htm
7. OpenAI Evals Adds Native Audio Input and Evaluation Features
OpenAI's Evals tool added native audio input and evaluation features, allowing developers to directly upload audio files for performance evaluation, significantly improving the development efficiency and accuracy of speech recognition and generation models.
AiBase Summary:
🎧 Native audio input functionality simplifies the evaluation process, improving development efficiency.
🔍 No need for text transcription to directly evaluate the performance of speech recognition and generation models.
💡 New features provide more accurate testing support for smart voice assistants and audio content generation.
8. Disrupting Tradition! Mini-o3 Open Source Model Achieves Ultra-Long Visual Reasoning, Deep Thinking Is No Longer a Problem
Mini-o3 is an open-source visual reasoning model jointly launched by ByteDance and the University of Hong Kong, capable of performing dozens of rounds of visual reasoning, significantly improving the ability to handle complex visual problems. Its core design includes the VisualProbe dataset, iterative data collection process, and ultra-round mask strategy, providing a new direction for multi-round visual reasoning technology.
AiBase Summary:
🧠 Mini-o3 achieved dozens of rounds of visual reasoning capabilities, breaking through the previous limit of 1-2 rounds of dialogue.
📊 By building the VisualProbe dataset and iterative data collection process, it improved the model's deep reasoning capabilities.
🔄 The ultra-round mask strategy optimized training efficiency, making the model perform better during testing.
Details link: https://arxiv.org/pdf/2509.07969
9. Shanghai AI Lab Launches Lumina-DiMOO, Pioneering a New Era of Multimodal Generation and Understanding
Shanghai Artificial Intelligence Laboratory, in collaboration with multiple universities, launched the next-generation multimodal generation and understanding model Lumina-DiMOO. The model adopts an innovative fully discrete diffusion architecture, effectively integrating and aligning text, images, and audio data through contrastive learning technology, significantly improving generation quality and efficiency, and showing broad application potential in various scenarios.
AiBase Summary:
🌟 Lumina-DiMOO is a new generation of multimodal generation model that adopts an innovative "fully discrete diffusion architecture" to improve data processing efficiency.
🛠️ The model achieves effective alignment and understanding of text, images, and other data through contrastive learning technology.
🚀 Lumina-DiMOO performs excellently in image generation and understanding, able to adapt to various application scenarios, showing broad application potential.
Details link: https://github.com/Alpha-VLLM/Lumina-DiMOO
10. Tencent's New AI Painting Upgrade! Fine-tuning Technology Increases Image Beauty by 300%
Tencent's fine-tuning technology significantly enhances the realism and aesthetic score of AI-generated images. Its innovative methods include "Direct-Align" and "Semantic Relative Preference Optimization," effectively solving the issues of reward cheating and offline adjustment limitations.
AiBase Summary:
🧠 "Direct-Align" technology reduces gradient explosion, improving model optimization capabilities.
🎨 "Semantic Relative Preference Optimization" (SRPO) enables text control over image style adjustments.
📈 Experiments show that SRPO-trained models significantly improve in realism and aesthetic quality.
Details link: https://arxiv.org/pdf/2509.06942
11. Meta AI Releases MobileLLM-R1: Lightweight Edge Inference Model, Parameters Less Than 1 Billion, Performance Significantly Improved
Meta AI's MobileLLM-R1 series model performs well in lightweight and edge computing fields, with parameter sizes ranging from 140M to 950M, focusing on mathematics, coding, and scientific reasoning. The model outperforms similar models in training efficiency and performance, especially in mathematics and coding tasks.
AiBase Summary:
🧩 New model released: Meta AI launches the lightweight edge inference model MobileLLM-R1 series, with parameters ranging from 140M to 950M.
📊 Training efficiency: MobileLLM-R1 uses only about 11.7% of the data for training, performing well, significantly reducing training costs and resource requirements.
💡 Performance advantage: In multiple benchmark tests, MobileLLM-R1-950M outperforms several large open-source models, especially in math and coding tasks.
Details link: https://huggingface.co/facebook/MobileLLM-R1-950M
12. Tencent Launches AI Application Prosperity Plan, Over 300 Enterprises Compete for the Intelligent Body New Track
Tencent's Global Digital Ecosystem Conference announced the AI Application Prosperity Plan, focusing on vertical scenarios to promote the deep penetration of AI in industries. The plan includes the AI Co-Creation Camp and AI Hundred Schools Campaign, attracting over 300 enterprises to participate, and fostering intelligent bodies and large model applications through technical sharing and resource opening.
AiBase Summary:
Tencent AI Application Prosperity Plan aims to deeply integrate AI into vertical scenarios, with two core modules: AI Co-Creation Camp and AI Hundred Schools Campaign.
The first offline event has attracted nearly 3,000 participants from multiple industries, showing strong market demand for large-scale AI applications.
Tencent provides technical support, resources, and content to help partners commercialize their AI solutions.
13. Google DeepMind Releases VaultGemma with Differential Privacy Capabilities
Google DeepMind's VaultGemma is a language model with differential privacy capabilities, focusing on protecting user data privacy. It is based on the Gemma2 architecture, uses a multi-query attention mechanism, and adds random noise to ensure that model outputs cannot be associated with specific training samples. Although its performance is slightly conservative, VaultGemma provides stronger protection for privacy and is expected to offer users a safer and more reliable experience in the future.
AiBase Summary:
🔒 VaultGemma is an open-source language model with differential privacy capabilities, with a parameter scale of 1 billion.
🧠 It uses a decoder-only Transformer design, with a sequence length limit of 1024 tokens.
🌐 Google will publicly release VaultGemma and its code library on Hugging Face and Kaggle, promoting the combination of privacy security and open-source technology.
14. QuestMobile Data: Doubao Surpasses DeepSeek, Ranking First in China's Native AI APP
QuestMobile's August 2025 AI application industry monthly report showed that Doubao, with a 6.6% month-on-month growth rate, reached 157 million monthly active users, surpassing DeepSeek to become the top native application. Tencent Yuanbao also performed well, with a monthly active user growth rate of 22.4%, ranking third among native applications. In addition, more than half of the top 50 AI applications are In-App plugin applications, and Doubao, as a PC client application, successfully entered the list, demonstrating its cross-end usage advantages.
