Chinese AI company DeepSeek has released an upgraded version of its flagship open-weight model, DeepSeek-R1-0528, which demonstrates notable improvements in reasoning, mathematics, programming, and logic. This model is reported to rival the performance of closed AI models such as OpenAI's o3 and Google's Gemini-2.5 Pro across several benchmarks. An 8-billion parameter distilled variant, DeepSeek-R1-0528-Qwen3-8B, outperforms Alibaba's Qwen3-8B model by one IQ point, achieves a 52 IQ score, and runs efficiently on a single 16GB GPU. There are claims that DeepSeek's latest model may have been trained using data from Google's Gemini AI, raising concerns among developers. Meanwhile, Alibaba has unveiled new open-source AI embedding models, maintaining its global leadership in text-embedding services as recognized by Hugging Face benchmarks. Alibaba's Qwen models are acknowledged by Nvidia CEO Jensen Huang as among the best open-source AI models, contributing positively to U.S. AI development. The release of DeepSeek's and Alibaba's models highlights China's growing leadership in open-source AI, with numerous startups emerging to develop AI agents that automate tasks ranging from app development to travel planning. Additionally, OpenThinker3, a 7-billion parameter model trained solely with supervised fine-tuning and no reinforcement learning, has been introduced, outperforming other open 7B and 8B models in math, code, and science. This surge in Chinese AI innovation follows the popularity of Butterfly Effect's Manus AI agent, which has catalyzed a boom in AI agent development within China despite local internet restrictions.