In the past 24 hours (November 27 to November 28, 2025), the global AI field has seen intensive new trends in basic models, video and music generation applications, industrial policies, and computing power infrastructure. Many domestic manufacturers are focusing on mathematical reasoning, video/music generation and e-commerce agents, while overseas are playing games around AI regulatory rules, data center heat dissipation and computing power financing risks.
1. DeepSeek released DeepSeekMath-V2, the first open source model to reach the IMO gold medal level
DeepSeek announced that its mathematical reasoning model, DeepSeekMath-V2, has solved 5 out of 6 problems in the 2025 International Mathematical Olympiad, reaching the gold medal line, and is widely regarded as the first open-source model to reach this level. The model weights are based on the Apache 2.0 protocol, which highlights the potential of the open source community to catch up and surpass in difficult reasoning compared to Google and OpenAI's previous closed-source experimental models.
2. Weibo's open source small parameter mathematics model VibeThinker challenges "scale is intelligence" with 1.5 billion parameters
Weibo announced the first open-source model, VibeThinker, with only about 1.5 billion parameters and a post-training cost of about $7,800, but it achieved results close to super-large models in difficult mathematical tests such as AIME2024/2025 and HMMT2025. Through reinforcement learning algorithms and fine task design, the team allows the "small model" to approach hundreds of billions of parameter-level opponents in mathematical reasoning, providing a powerful sample for low-cost and high-intelligence routes.
3. Zhipu Qingying 2.0 is online: text directly out of 1080P video and automatic sound effects
Zhipu AI launched Qingying 2.0, which supports the generation of 1080P high-definition short videos with a sentence of up to about 10 seconds, and combines the self-developed CogSound model to automatically match ambient sound and action sound effects to achieve an "audio-visual integration" generation experience. The new version is fully upgraded in motion camera control, multi-style graphics and inference costs, and is positioned as a video production infrastructure for industries such as finance, e-commerce, advertising and film and television.
4. Kunlun Wanwei Mureka V7.6/O2 was released to accelerate the commercialization of AI music
Kunlun Wanwei announced that the AI music platform Mureka has been upgraded to V7.6 and launched the O2 model, which continues to strengthen the arrangement structure, timbre details and style consistency. The new version has significantly improved speed and stability, supporting larger-scale commercial music generation scenarios for global users, and the official emphasized that its goal is to become a music AI base that can "create, collaborate, and interact in real time".
5. 1688 released the cross-border AI agent "Ao Shrimp", opening up product selection to factory docking with one click
Alibaba's 1688 launched the cross-border e-commerce AI agent "Ao Shrimp" (AlphaShop), which integrates image recognition, link analysis and natural language interaction, providing overseas merchants with one-stop services from market research, product selection analysis to factory sourcing and intelligent inquiry. Merchants only need to upload screenshots or product links to automatically match factories and generate marketing content, reflecting the evolution of e-commerce platforms to "intelligent supply chain operating systems".
6. Ministry of Industry and Information Technology: AI has become a "catalyst" to promote consumption, and 515 million people have used generative AI products
TheMinistry of Industry and Information Technology said at the regular policy briefing of the State Council that it will make efforts from both product innovation and scenario innovation to accelerate the implementation of artificial intelligence in the field of consumer goods. According to the official disclosure, as of the first half of this year, the number of users of domestic generative AI products has reached about 515 million, and it will focus on promoting scenarios such as smart home appliances, whole-house intelligence and "artificial intelligence + health care" to release the demand for large models for terminal consumption.
7. Meta open-source CoT-Verifier, making large model inference errors "locable and correctable" Meta
launched CoT-Verifier based on Llama 3.1, which transforms complex reasoning processes into analyzable "graph structures", and then uses lightweight classifiers to identify wrong steps to realize automatic diagnosis of white-box inference chains. The official plan is to extend this set of "graph intervention" methods to code generation and multimodal reasoning, which is expected to significantly reduce the hallucination rate in long-chain inference and provide a new technical path for "reliable large models".
8. Google tightens Gemini 3 Pro's free permissions, and the free user quota is dynamically adjusted
Google recently adjusted its free usage strategy for Gemini 3 Pro and Nano Banana Pro, canceling the previously stable free call limit and instead dynamically adjusting the number of daily available calls according to capacity pressure, and reducing the free image generation quota. At the same time, some new features of NotebookLM are only open to paying users, indicating that leading AI manufacturers are accelerating from "pulling new for free" to a more refined subscription monetization model in the context of high computing power costs.
9. The European Union launches a simplification plan for digitalization and AI rules, claiming to save more than 155 billion euros in costs for enterprises
TheEuropean Commission announced a new digital reform plan, which plans to reduce duplicate filings and cross-border compliance costs by harmonizing and streamlining data and AI-related compliance requirements. Official estimates suggest that it is expected to save about 155 billion euros in the medium to long term, aiming to reduce the burden on European AI and data-driven enterprises without weakening security and privacy, and enhance the attractiveness of competition with the United States and China in the AI industry.
10. The computing power ecosystem around OpenAI is burdened with about $96 billion in debt, raising concerns about financial stability
The latest analysis shows that in order to meet OpenAI's huge demand for computing power and data centers, its core infrastructure partners have accumulated about $96 billion in debt to build computer rooms, purchase GPUs, and expand power capacity. At the same time, OpenAI's own long-term energy and computing power contracts are worth trillions of dollars, and the current revenue scale is still limited, and the outside world is worried that if the pace of AI commercialization falls short of expectations, the related debt chain may amplify financial risks.
11. AI has ignited the expansion of global data centers, and heat dissipation and energy consumption challenges have erupted
Recently, CME Group has interrupted transactions due to the failure of the cooling system of cooperative data centers, once again highlighting the heat dissipation risk of high-density AI data centers. The industry expects that up to 40% of data center energy consumption is used for cooling, and new solutions such as liquid cooling and water cooling are accelerating the replacement of traditional air cooling, while also bringing new problems such as water consumption and operation and maintenance complexity. How to control energy consumption and failure risks while supporting the explosion of AI computing power is becoming an important test question for infrastructure investment.
Frequently Asked Questions (Q&A)
Q: What is the most noteworthy thing about large models and basic research directions in the past 24 hours?
A: First, DeepSeekMath-V2 released by DeepSeek has reached the gold medal level in the simulation IMO and is completely open source, marking the first time that the open source community has benchmarked or even approached the closed-source giant in terms of the difficulty of the top mathematical competition. Second, Weibo's small-parameter model VibeThinker uses 1.5 billion parameters to achieve high scores on multiple mathematical benchmarks, proving that through more refined training paradigms, small models can also be competitive in difficult reasoning.
Q: What are the obvious opportunity signals for domestic developers and entrepreneurs in this issue?
A: From the Qingying 2.0 and Mureka upgrades, it can be seen that video and music AIGC is rapidly moving from "fun" to "commercial", and tools and SaaS opportunities around short video production, advertising creativity, and music copyright materials are increasing. 1688 "Ao Shrimp" shows that there is a strong demand for agents in cross-border e-commerce scenarios, and "industry agents" who understand supply chains and vertical industry processes will become more and more valuable.
Q: What are the implications of new developments in overseas regulation and infrastructure?
A: The EU is trying to reduce compliance costs for enterprises by unifying and simplifying AI and data compliance requirements, which is conducive to the landing of small and medium-sized AI companies in the European market. At the same time, the focus on computing debt and data center cooling issues surrounding OpenAI has also reminded the entire industry that it must find a more sustainable balance between business models, energy consumption and financial risks.
Q: What does the tightening of free quotas mean for ordinary users and small and medium-sized enterprises?
A: Google has adjusted the free strategy of Gemini 3 Pro, indicating that the cost pressure of "high-quality large model + large computing power" is difficult to fully digest with free products. While individual users may need to switch between multiple models to control costs, SMEs need to evaluate which tasks must rely on top-level closed-source models and which can be migrated to open source or dedicated small models to reduce long-term subscriptions and computing power spending.
