Open source models (OSS) have emerged as a key growth engine in the AI ecosystem, according to the "State of AI" report by OpenRouter and a16z. The report, based on over 100 trillion tokens of LLM usage data from the past year, highlights that OSS usage has surged to account for about one-third of total token volume by the end of 2025. Notably, Chinese OSS models represent 13.0% of this usage, with the rest of the world contributing 13.7%, while proprietary models still dominate with 70%. The report identifies significant trends, including the rapid adoption of new models like DeepSeek V3, Kimi K2, and Qwen 3 Coder, which have quickly gained traction post-launch. The top ten OSS models processed over 33 trillion tokens, led by DeepSeek (14.37 trillion), Qwen (5.59 trillion), and Meta LLaMA (3.96 trillion). Medium-sized models (15-70 billion parameters) are gaining popularity, shifting usage from smaller to larger models. OSS models are primarily used for role-playing and programming tasks, with over half of open-source usage dedicated to narrative and emotional interactions, indicating potential for consumer applications. Chinese AI models, including Qwen and Moonshot AI's Kimi, have significantly contributed to global OSS growth, showcasing technical competitiveness. Chinese OSS models have increased their weekly token share from 1.2% at the end of 2024 to nearly 30% by the end of 2025. These models excel in multilingual support and cultural adaptability, marking a shift towards decentralized global AI and highlighting the importance of multilingual capabilities over sheer scale.