Whenever I write long articles, no one reads them. I'm preparing to write a few in-depth pieces to submit. 🥹

Look at OpenRouter @OpenRouterAI and a16z @a16z report "State of AI" This report is based on real LLM usage data of over 1 trillion tokens in the past year. I have organized the sections on open source and Chinese AI. # Open Source Models Open source models (OSS) are viewed in the report as the core growth engine of the AI ecosystem, with usage rapidly climbing from a low share at the beginning of the year to about one-third of the total token volume by the end of 2025. Key statistics and trends include: * OSS models account for an average of 30% of total usage, with Chinese OSS at 13.0%, and OSS from other regions at 13.7%, while proprietary models still dominate at 70%. The peak in growth coincides with the release of new models, such as DeepSeek V3, DeepSeek R1, Kimi K2, the GPT OSS series, and Qwen 3 Coder, which saw rapid adoption within weeks of their release. * In terms of token volume, the top ten OSS models collectively processed over 33 trillion tokens, with DeepSeek (14.37 trillion), Qwen (5.59 trillion), and Meta LLaMA (3.96 trillion) leading the way. * Medium-sized models (1.5-7 billion parameters) have become an emerging hotspot (e.g., Qwen2.5 Coder 32B), with usage shifting from small models to medium and large models. * OSS models play a dominant role in role-playing (accounting for 52% of OSS tokens) and programming tasks. The report notes that over half of the open source usage is for narrative and emotional interactions, suggesting a huge opportunity for consumer applications. * OSS models focus on low-cost, high-volume tasks (e.g., cache optimization reduces actual costs below the listed price), while closed-source models handle high-value work. Price elasticity is increasing, driving the market from differentiation to liquidity. # Chinese AI Section Chinese open source models, including Qwen, Kimi from Moonshot AI, and DeepSeek, have driven global OSS growth and demonstrated competitiveness in the tech field. Key statistics and trends include: * Usage share: Chinese OSS averages 13.0% of weekly token volume, growing from 1.2% at the end of 2024 to nearly 30% in some weeks by the end of 2025. * Chinese models are known for rapid iteration, with regular updates from DeepSeek and Qwen reshaping the OSS landscape. Top Chinese OSS contributors include DeepSeek (14.37 trillion tokens), Qwen (5.59 trillion), Minimax (1.26 trillion), and MoonshotAI (0.92 trillion). Inference models like Kimi K2 have quickly captured market share due to cost efficiency and flexibility. * Chinese OSS accounts for 33% in role-playing, but programming and technical tasks have a higher share (39%, above the overall OSS average of 38%), showing its advantage in productivity tools. Role-playing traffic is expected to balance out by the end of the year (RoW OSS 43%, Chinese OSS ~8% early on, then growing). * Chinese models excel in multilingual support (e.g., translation), benefiting from cultural adaptability and customization. The report emphasizes that the rise of Chinese AI marks a decentralization of global AI, with the success of non-Western OSS indicating that competition will focus on multilingual capabilities and cultural adaptability rather than sheer scale.