LLM Model Technology Updates
Google has released the TimesFM-2.5 time series forecasting model, featuring 200 million parameters and a single-decoder architecture. This represents a significant reduction from the previous generation’s 500 million parameters and multi-decoder design. TimesFM-2.5 supports a context length of up to 16,000 data points and enables local probability predictions. Its GIFT-Eval accuracy leads its peers. The model is applicable in retail demand forecasting, weather monitoring, supply chain optimization, and more. It is now available on Hugging Face and will be integrated into BigQuery and ModelGarden, promoting the wide adoption of zero-shot time series forecasting in practical applications.

Tencent AI Lab, in collaboration with universities, has developed the Paralel-R1 reinforcement learning framework to address the weak generalization of supervised fine-tuning (SFT) models. By generating “parallel thinking data” using simple prompts and training with a progressive curriculum, combined with an alternating reward strategy prioritizing accuracy and supplemented by thought usage rewards, the framework has improved average accuracy on math benchmarks by 8.4% and boosted AIME25 performance by 42.9%.

AI Industry and Market News
Google has launched a new AI shopping payment protocol called Agent Payments Protocol (AP2). This protocol enables AI agents to autonomously complete transactions on behalf of users after authorization, eliminating the need for repeated manual approvals. It includes anti-fraud mechanisms and traceable audit functions. Users can set “intent authorization” to specify their needs and budgets, allowing for manual approvals online or automated offline purchases—for example, pre-setting conditions for the agent to purchase concert tickets at scheduled times.

Disney, Warner Bros., and Universal Pictures have jointly filed a lawsuit against Chinese AI company MiniMax, accusing it of illegally using characters from franchises like Star Wars for commercial promotion via its “Conch A!” product. The studios seek damages based on the profits earned and statutory compensation (up to $150,000 per copyright), as well as a permanent injunction. The outcome of this lawsuit could have significant implications for copyright regulations within the AI industry.

On September 17, Hello TransTech announced strategic investment from Alibaba in its Robotaxi business. The two companies will deepen cooperation in intelligent driving large models and related fields. Hello TransTech entered the autonomous driving sector in June this year, partnering with Ant Group and CATL to establish a company focused on L4 autonomous driving research and development, with initial funding exceeding 3 billion RMB and a technical team of over 200 people.