Today, I took a deep dive into the latest news in the AI sector and found several interesting developments. Anthropic has upgraded Claude Opus to version 4.7 with enhanced coding and agent capabilities; ZhiYuan held a 2,500-person partner conference showcasing four new models, indicating rapid progress in this field; Alibaba’s world model has emerged, achieving the top global ranking in video editing. On the digital front, China’s average daily token usage has surpassed 140 trillion, and a Stanford report indicates that the gap between Chinese and American models has narrowed to just 2.7%. Here are 15 noteworthy updates from the last 24 hours.

1. Anthropic Officially Releases Claude Opus 4.7: Comprehensive Upgrades in Coding and Agent Capabilities
   Key Points: Anthropic has officially launched Claude Opus 4.7, its latest flagship model. The new version shows significant improvements in coding, agent tasks, visual understanding, and multi-step tasks, described as “stronger performance and higher consistency”. Anthropic’s current valuation has reached $80 billion. Notably, after the launch of Opus 4.7, Opus 4.5 was removed from the store, making Opus 4.7 the flagship model.
   Commentary: Each flagship update from Anthropic raises the bar for coding capabilities, and this time, version 4.7 focuses on enhancing the consistency of agent tasks—meaning the AI will not easily “get lost” or be inconsistent during multi-step tasks. This is a critical improvement for scenarios requiring long-term autonomous work.

2. Claude Code v2.1.111 Released: New xhigh Effort Level and Auto Mode Fully Opened
   Key Points: Claude Code has released two versions, v2.1.111 and v2.1.112. A new xhigh effort level (between high and max) has been introduced along with an interactive effort slider (/effort command), allowing users to control the AI’s depth of thought more flexibly. The Auto mode is now fully available to Max subscribers, eliminating the need to wait for beta testing qualifications.
   Commentary: The xhigh tier is a clever design—high is too shallow, max is too expensive and slow, and xhigh fits perfectly in the middle. The full opening of Auto mode further lowers the entry barrier for Claude Code, which is great news for individual developers.

3. Google MaxText Adds SFT and RL Support on Single-Host TPU
   Key Points: Google’s MaxText framework has added support for supervised fine-tuning (SFT) and reinforcement learning (RL) training on single-host TPU configurations, using JAX and Tunix libraries for model fine-tuning. This update significantly lowers the barrier for post-training models on Google TPU—previously requiring large clusters, now a single TPU machine can run it.
   Commentary: Google has been filling gaps in its TPU ecosystem toolchain. Previously, post-training on TPU was a major pain point, and this MaxText update greatly enhances TPU usability, making it practical for teams with limited budgets looking to fine-tune using TPUs.

4. NVIDIA Open-Sources Quantum AI Model Ising: Reduces Calibration Time from Days to Hours
   Key Points: NVIDIA has released the open-source quantum AI model “Ising”, designed for calibration optimization of quantum computers. Based on the statistical physics Ising model framework, it reduces the calibration time of quantum systems from several days to just hours. NVIDIA states this is a crucial part of its quantum computing ecosystem strategy.
   Commentary: Quantum computing has gained significant attention in recent years, but calibration is notoriously time-consuming—taking days to make a quantum computer usable. Ising’s reduction of calibration time to hours is a significant efficiency improvement for the practical application of quantum computing. Although large-scale commercial use is still a distance away, NVIDIA’s investment in infrastructure is noteworthy.

5. Anthropic’s MCP Protocol Found to Have Design Flaws: Over 200,000 AI Servers at Risk
   Key Points: Security researchers disclosed that Anthropic’s MCP (Model Context Protocol) has design flaws that could lead to remote code execution vulnerabilities, affecting over 200,000 AI servers worldwide. The MCP protocol is a standard promoted by Anthropic for model context, widely adopted for data integration in AI applications. Details of the vulnerability have been submitted to Anthropic, which is currently working on a fix.
   Commentary: The emergence of a security flaw shortly after the MCP protocol’s promotion is significant. While 200,000 affected servers is a large number, the issue lies at the protocol design level, and once patched, it shouldn’t leave a lasting backdoor. Applications using MCP should stay updated with official patches during this period.

6. Alibaba Releases World Model HappyOyster: Global Ranking First in Video Editing
   Key Points: Alibaba’s ATH Innovation Division team has released the next-generation open-world model HappyOyster, supporting real-time interaction and multi-modal creation. HappyHorse-1.0 has achieved the top global ranking in video editing. The model is based on a native multi-modal architecture, supporting multi-modal understanding and audio-video joint generation, and belongs to the same team as the previous HappyHorse.
   Commentary: The Alibaba ATH team has been making continuous strides, with HappyHorse followed by HappyOyster, which has reached the top rank in video editing. This team’s product iteration speed is impressive, and they have a clear goal—to establish a technological barrier in the multi-modal generation direction. Future open-source developments are worth watching.

7. Tencent Open-Sources Huanyuan 3D World Model 2.0: One-Click Generation of Editable 3D Assets
   Key Points: Tencent has officially open-sourced Huanyuan 3D World Model 2.0, which supports one-click generation of editable 3D assets that can be exported in various formats and seamlessly integrated with mainstream game engines like Unity and Unreal. This is a significant update in Tencent’s 3D generation direction, showing clear improvements in generation quality and editing flexibility compared to version 1.0.
   Commentary: The demand for 3D content in gaming and virtual reality is enormous, and Tencent’s open-sourcing of Huanyuan 2.0 effectively connects with mainstream game engines, creating a complete workflow. This news is more practical for game developers than many large model releases.

8. ZhiYuan Robotics 2026 Partner Conference: Four New Models and Four Major AI Models Unveiled
   Key Points: ZhiYuan Robotics held its 2026 Partner Conference in Shanghai, attracting 2,500 partners from 34 countries and regions. At the conference, ZhiYuan unveiled four new robot models, four AI models, seven solutions, and open datasets. The theme of the conference was “Opening the Era of Scalable General Robots”.
   Commentary: The participation of 2,500 people from 34 countries indicates ZhiYuan’s significant progress in internationalization and ecosystem development. The simultaneous release of four new models suggests an acceleration in mass production, while the combination of four major models indicates simultaneous advancements in their “brain” and “small brain”. The competition among leading players in the embodied intelligence sector has entered a new phase.

9. Ant Group Open-Sources LingBot-Map: Real-Time 3D Reconstruction with a Single Camera
   Key Points: Ant Group has open-sourced the robot navigation framework LingBot-Map, which supports real-time 3D reconstruction using a single camera, specifically designed for robot navigation and autonomous driving scenarios. This framework can complete complex environment 3D modeling relying solely on a single camera, lowering the hardware barrier for robot navigation.
   Commentary: Single-camera 3D reconstruction is a challenging task—without depth sensors, it relies purely on visual algorithms to infer 3D structures. Ant’s ability to open-source this indicates that they have successfully implemented it internally. This is beneficial for low-cost embodied intelligence hardware and allows more small to medium teams to participate in robot navigation development.

10. DeepSeek V4 Planned for Release by End of April: Trillion Parameter Scale, Deep Compatibility with Domestic Chips
    Key Points: DeepSeek founder Liang Wenfeng revealed that the next-generation flagship model DeepSeek V4 is scheduled for official release by the end of April. V4 will feature a trillion parameter scale and deep compatibility with domestic chips. Liang also mentioned that DeepSeek has recently completed a new round of financing, but the specific amount has not been disclosed.
    Commentary: Liang’s mention of “trillion parameters + compatibility with domestic chips” is an important signal—indicating that DeepSeek V4 is not only a performance sprint but also has a clear domestic adaptation goal. If it is released on schedule by the end of April, it will mark a significant milestone for domestic large models at the trillion parameter level.

11. China’s Average Daily Token Usage Surpasses 140 Trillion: Year-on-Year Growth Exceeds 40%
    Key Points: According to Mao Shengyong, Deputy Director of the National Bureau of Statistics, at a press conference, as of March 2026, China’s average daily AI token usage has surpassed 140 trillion, a year-on-year increase of over 40%. This data marks a significant breakthrough in the commercialization and large-scale application of artificial intelligence in China.
    Commentary: A daily usage of 140 trillion tokens is a substantial figure, and a 40% growth rate indicates that enterprises are accelerating their investment in AI applications. However, it is important to note that token usage reflects volume, not directly equating to commercial value—it still depends on the actual business output generated from these calls. Nonetheless, the penetration rate of AI in Chinese enterprises is significantly increasing.

12. Stanford’s 2026 AI Index Report: Gap Between Top Models in China and the US Only 2.7%
    Key Points: Stanford University’s Human-Centered AI Institute (Stanford HAI) released the 2026 AI Index Report. The report shows that China holds advantages in several AI metrics, with the performance gap between top models in China and the US shrinking from 8% last year to 2.7%. Alibaba ranks third in the global top model contribution list for 2025, while also being the Chinese tech company with the most selected important models.
    Commentary: A 2.7% gap is nearly negligible in technical terms—indicating that top models in China are now on par with the strongest models in the US. Alibaba’s entry into the global top three is a significant event, suggesting that Chinese tech companies are beginning to see substantial returns on their investments in foundational model development.

13. Liang Wenfeng Reportedly Secures 2 Billion Yuan in Financing
    Key Points: DeepSeek founder Liang Wenfeng has reportedly completed a new round of financing amounting to approximately 2 billion yuan. This round of financing will primarily be used for the research and development of DeepSeek V4 and the procurement of computing resources. DeepSeek has not yet commented on the financing news.
    Commentary: If the figure of 2 billion yuan is accurate, this would be one of the most significant single financings in the large model sector this year. DeepSeek has maintained a relatively low profile in its commercialization path, but its technology and talent density are widely recognized. How this financing will be utilized and whether V4 can open up a market in commercial settings is worth ongoing attention.

14. THINKINGAI Launches Enterprise-Level Agent Platform AGENTIC ENGINE
    Key Points: THINKINGAI officially launched the enterprise-level agent platform AGENTIC ENGINE in Silicon Valley, designed for the development and deployment of AI agents in enterprise scenarios. The platform supports multi-model collaboration, workflow orchestration, and enterprise-grade security controls, having already secured early adoption agreements with several Fortune 500 companies.
    Commentary: Enterprise-level agent platforms are currently very popular, with Anthropic and Cohere also pushing similar solutions. AGENTIC ENGINE’s debut in Silicon Valley clearly targets international large enterprises. Whether it can capture market share in an already crowded field will depend on product experience and pricing.

15. ELEPHANT Gains Popularity: AI Begins to Calculate “Token Waste”
    Key Points: The AI tool ELEPHANT has recently gained popularity in the developer community, specifically designed to help enterprises analyze and optimize token consumption in AI model calls, identifying areas of severe waste and providing optimization suggestions. As the cost pressure of large model APIs increases, the demand for such “token auditing” tools is rapidly growing.
    Commentary: Token costs are one of the core expenses in large model applications, and as business scales grow, the value of token optimization becomes evident—saving tokens translates to pure profit. The rise of tools like ELEPHANT indicates that large model applications are transitioning from “just running” to “refined operations”, signaling industry maturity.