Welcome to the 【AI Daily】 section! This is your guide to exploring the world of artificial intelligence every day. Each day, we bring you the latest hot topics in the AI field, focusing on developers and helping you gain insights into technological trends and innovative AI product applications.
Fresh AI products click to learn more: https://top.aibase.com/
1. Moon's Dark Side releases new open-source model Kimi-Dev-72B, breaking programming benchmark records
The Moon's Dark Side has launched an open-source model, Kimi-Dev-72B, which focuses on software engineering tasks. It achieved the highest score of 60.4% in the SWE-bench Verified test, surpassing DeepSeek-R1 with its 72 billion parameters, showcasing its strong capabilities in programming.
【AiBase Summary:】
🚀 With only 72 billion parameters, Kimi-Dev-72B scored 60.4% in the SWE-bench Verified test, becoming a new benchmark for open-source models.
🔍 The model combines BugFixer and TestWriter roles to ensure code quality and correctness, improving performance through self-play mechanisms.
🌟 Future plans include deep integration with popular development tools, continuous optimization, and launching even stronger versions.
2. MiniMax-M1 is now open-source! 1M ultra-long context reasoning, AI new king built at a cost of only $530,000!
MiniMax-M1 has drawn attention due to its ultra-long context reasoning ability, efficient training costs, and open-source nature, standing out among open-source models.
【AiBase Summary:】
Context window reaches 1M input and 80k output, far exceeding GPT-4o, suitable for complex document analysis and multi-turn conversations.
Training cost is only $530,000, achieving efficient inference and low cost through MoE architecture and CISPO algorithm.
Open sourced on Hugging Face platform, supporting 40k and 80k thought budgets, performance comparable to top commercial models.
3. Tencent's LeVo arrives! An AI singing model comparable to Suno 4.5, supporting zero-shot voice cloning
Tencent AI Team’s LeVo model has sparked heated discussions with its powerful voice cloning, track generation, and high-fidelity music performance. Compared to Suno4.5, it performs excellently in multiple key metrics while supporting zero-shot voice cloning and track generation, showcasing revolutionary breakthroughs.
【AiBase Summary:】
🌟 Supports zero-shot voice cloning, accurately replicating voices with just 3 seconds of audio, significantly lowering the threshold for music creation.
🎵 Provides track generation mode, enabling separation of vocals and accompaniment, offering higher flexibility for professional music production.
🌐 Released in open source format, promoting global music creation communities and enhancing China's international influence in AI technology.
Details link: https://levo-demo.github.io/
4. Alibaba releases Qwen3 upgrade version compatible with Apple MLX architecture
I am very optimistic about this cooperation between Alibaba and Apple. The Qwen3 upgrade not only supports more languages but also enhances performance and reasoning ability. This marks an important step forward in Apple's intelligence in the Chinese market.
【AiBase Summary:】
🌟 Alibaba launches Qwen3 upgrade, compatible with Apple MLX architecture, assisting Apple's intelligence in China.
📱 New Qwen3 supports 119 languages with stronger performance and hybrid reasoning ability.
🚀 Apple's intelligence has not yet been launched in China and may be previewed in iOS18.6 official public beta version.
5. BeanPod adds "AI Podcast" function to PC and web versions
BeanPod introduces a new 'AI Podcast' feature, generating natural two-person dialogue podcasts by uploading PDFs or links, revolutionizing information reception methods.
【AiBase Summary:】
🌟 Quickly generate natural and fluent two-person dialogue podcasts just by uploading PDFs or links.
🏃♂️ Suitable for work, learning, and various scenarios, efficiently acquiring information during fragmented time.
🎙️ Realistic voice effects, eliminating machine-like feel, providing immersive auditory experience.
6. Quark App launches "Quark Teacher" with personalized AI tutoring capabilities
Quark App has introduced a new learning product called 'Quark Teacher'. This AI tutor is powerful, capable of explaining problems, grading homework, creating questions, and finding past exam papers. It excels particularly in solving math and physics problems, and has the ability to teach according to individual student needs, providing personalized tutoring by analyzing students' learning data.
【AiBase Summary:】
✨ Integrates multiple learning functions such as problem explanation, homework grading, question creation, and exam paper searching, supporting in-depth analysis of math and physics problems.
🎯 Can provide personalized tutoring based on student characteristics, simulating real teacher teaching ideas to help students understand and improve learning outcomes.
📚 Has massive question bank resources, including professional question banks and school exam papers, meeting diverse learning needs.
7. Panasonic's new OmniFlow multimodal large model enables free switching between text, image, and audio
I am very optimistic about OmniFlow, this multimodal large model. Not only can it easily convert text, images, and audio, but it also allows users to customize the generated results according to their needs, greatly enhancing operational flexibility and efficiency.
【AiBase Summary:】
🌟 OmniFlow supports efficient conversion between text, images, and audio, bringing a new multimodal experience.
⚙️ Uses modular design, each component is independently pre-trained, improving resource utilization efficiency and optimizing training effects.
🎯 Introduces multimodal guidance mechanisms, allowing users to precisely control the generation process to meet diversified needs.
8. TikTok's new Symphony AI tool goes live: images turn into videos, text directly generates ads
TikTok has launched three AI video creation tools, including 'Image to Video', 'Text to Video', and 'Showcase Products', aiming to simplify the brand advertising content production process. These tools are integrated into Symphony Creative Studio and cooperate with Adobe Express and WPP Open to enhance advertiser efficiency.
【AiBase Summary:】
✨ Image to Video function makes static images easily transform into dynamic videos, generating multiple AI video options just by uploading images and adding text prompts.
📝 Text to Video function does not require images or templates, making videos just by using text, helping advertisers quickly test and refine creativity.
🛍️ Showcase Products tool fuses product images with digital avatars, creating an immersive advertising experience, enhancing user-generated content styles.
9. ZEEKR collaborates with Volcano Engine, BeanPod large model empowers intelligent cabin new experiences
ZEEKR Automobile collaborates with Volcano Engine to integrate the BeanPod large model into the new version of ZEEKR AI OS, enhancing intelligent cabin service capabilities and optimizing personalized experiences.
【AiBase Summary:】
BeanPod large model integrated into ZEEKR intelligent cabin, achieving precise recommendations and personalized services.
Upgraded ZEEKR intelligent voice assistant Eva supports seamless switching from traditional voice interaction to large language model services.
ZEEKR's 500,000th model, 009, rolls off the assembly line, setting a new record for the fastest luxury pure electric vehicle production.
10. Major breakthrough in large models! Meta Llama 3.1 can recall 42% of the content from Harry Potter!
Research by Stanford University and other institutions shows that Meta's Llama3.170B model performs excellently in text memory, especially in popular books like Harry Potter.
【AiBase Summary:】
📚 Llama3.170B model can remember 42% of the content in Harry Potter, far exceeding Llama165B's 4.4%.
🔍 Research uses the Books3 database, testing model memory capability through marking paragraphs.
🌟 Memory effect in popular books is better, showing AI progress in understanding and processing text.
11. Grok Task function launches! Timed tracking of X trending topics, AI efficiency exceeds ChatGPT
xAI’s AI assistant Grok has launched a new Tasks timed task function, automating queries and external notifications to provide users with an efficient and convenient information retrieval experience.
【AiBase Summary:】
🌟 Supports various task frequencies, from immediate to long-term tracking, meeting diverse needs.
📧 Provides external notification functions, such as email delivery, proactively delivering results to users, enhancing usability.
🏆 SuperGrok users enjoy higher quotas and priority access to cutting-edge features like DeepSearch and Big Brain Mode.
12. Gemini 2.5 Pro is about to update the Deep Think function
As an AI assistant, I am very excited about the new Deep Think function in Gemini 2.5 Pro. This function not only enhances AI's reasoning ability in complex tasks but also makes significant improvements in user experience and security. The launch of Deep Think gives me a glimpse of the infinite possibilities of AI in professional fields.
【AiBase Summary:】
💎 The Deep Think function significantly improves performance in complex tasks through multithreaded reasoning, performing exceptionally well in mathematics, programming, and multimodal tasks.
🌐 Users can intuitively switch to Deep Think mode via the web UI, with the function gradually opening up to more users.
🔒 Before formal release, Google collects feedback through APIs and conducts security assessments to ensure the stability and data security of the function.
13. Google Maps receives a massive upgrade: new AI features bring smart reviews and fuel-efficient routes
Google Maps has undergone comprehensive upgrades through the introduction of generative AI technology, enhancing navigation, exploration, and personalized recommendation functions to provide users with a smarter and more efficient experience.
【AiBase Summary:】
🌍 Use generative AI search function to achieve accurate location queries through natural language.
🔍 Smart review analysis function automatically summarizes user reviews and answers specific questions about locations.
🌿 Introduces fuel-efficient route optimization function, recommending more environmentally friendly driving routes by combining multiple factors.