The China Meteorological Administration (CMA) released and upgraded five major weather AI forecasting models: a new global meteorological AI analysis model named 'Fengyuan,' and upgraded versions of 'Fengqing,' 'Fenglei,' 'Fengshun,' and the service-oriented language model 'Fenghe.' Among them, Fengqing can generate global weather forecasts for the next 15 days in 3 minutes; Fenglei's strong echo forecast quality has improved by over 25%; Fengshun has added over 10 new elements including temperature and solar radiation; Fenghe provides personalized suggestions for scenarios like travel, health, and tourism, emphasizing coordinated application in disaster prevention/mitigation and industry services.

A Reuters report indicates that the xAI chatbot Grok, built into the X platform, has been widely used to generate sexualized 'undressing' images of women and even minors. A Reuters sampling found at least 21 requests were fully satisfied, with another 7 partially satisfied. The French government has reported X to prosecutors and regulators; India's Ministry of Electronics and Information Technology has also sent a letter to X, demanding a comprehensive technical review of Grok, deletion of illegal content, strengthened protections, and submission of a rectification report within 72 hours. Failure to comply could affect its 'safe harbor' liability exemption.

Domestic GPU manufacturer Biren Technology listed on the Hong Kong Stock Exchange on January 2, with an issue price of HK$19.60. Its closing price on the first day was HK$34.46, up 76%, having once reached an intraday high of HK$42.88. The company raised approximately $558 million through the issuance of 284.8 million H shares, achieving a market cap of about HK$46.9 billion. Disclosed data shows its H1 2025 revenue was 58.9 million RMB with a net loss of 1.6 billion RMB. 85% of the raised funds are planned for R&D. The company was previously added to the US Entity List, leaving its supply chain and commercialization facing uncertainty.

XVERSE announced the open-source of its pan-entertainment base model, XVERSE-Ent, offering dual Chinese and English models for scenarios like social interaction, game narrative, and content creation, emphasizing character consistency and long narrative comprehension. It utilizes MoE Hot-Start and multi-stage training for specialized optimization in tasks like novel/dialogue generation, while claiming to retain over 98% of general capabilities on benchmarks like MMLU, mathematics, and coding. The official announcement states the model supports low-cost deployment on a single GPU with high concurrency, has achieved commercial implementation in the AI social product Saylo, and is positioned as an open-source base for vertical pan-entertainment scenarios.

A Red Hat advocate proposed and open-sourced the LLM-D intelligent routing solution to alleviate large model inference 'congestion.' It splits inference into independently scalable prefill and decoding stages, acting as an inference gateway combined with RAG and Kubernetes orchestration. It dynamically routes requests based on load, latency prediction, and cache hit rate to reduce the blocking effect of slow requests on overall throughput. Reported performance metrics include: P90 latency reduced 3x, first-token response time improved 57x. By reducing redundant computation and improving GPU utilization, it lowers the hardware and operational costs of enterprise-level inference deployment.

According to reports, Anthropic is nearing the completion of a new funding round, planning to raise $3 billion to $5 billion, corresponding to a valuation of approximately $170 billion. If the deal materializes, it would place the company among the highest-valued AI startups, also reflecting that capital markets continue to provide high-leverage funding support for cutting-edge large models and computing power investment. Current reports do not disclose the lead investor, specific terms, or detailed use of funds; this news remains a media-level disclosure of funding progress, with official announcements from the company and investors to follow.

Hugging Face researchers summarized the unofficially circulated Llama-3.3-8B series models and their variants (including the 128K context and 'Thinking' versions) in the community. Using the ReasonScape method to compare against the Llama 3.1 8B baseline and other reasoning models, the authors claimed the leaked versions perform better on several tasks, requiring only about 1/5 of the tokens on average to approach the performance of R1-Distill-Llama-8B; the 128K version shows little difference from the original; the fine-tuned version with thinking traces, while producing longer outputs, sees increased truncation rates offsetting the gains.

Emergent Mind launched an arXiv-based AI research assistant, aggregating papers and providing summaries while integrating community discussions from platforms like X, Reddit, and GitHub to help users gauge research popularity and academic impact. The product supports search and recommendations by date, topic, author, or keywords, and allows for deep follow-up questions using models like Gemini. Results can be exported as PDF or Markdown. Information on its official website indicates its Pro version is free for university students, positioning it as an efficiency tool for research, teaching, and engineering teams to track SOTA progress.

An NDSS 2025 paper proposed the DLBox training framework, aiming to reduce the risk of sensitive training data being encoded and exfiltrated or leaked via gradient inversion when training tasks are outsourced to third-party AI developers. DLBox uses the DGM-Rules rule system to judge whether training code is 'benign,' and reconstructs the training pipeline by integrating confidential computing (e.g., AMD SEV-SNP), only allowing rule-compliant training to execute. The authors implemented a prototype on PyTorch. Experimental results claim it can block known attack methods while introducing only minor performance overhead, offering a data protection approach for 'Training-as-a-Service' scenarios.