The current balance of power in open models

Chinese open-weight models have become the main force in the open-model ecosystem, leading in downloads, benchmark performance, and usage on platforms such as OpenRouter and OpenCode. The article argues that American open models still have important adoption pockets, especially in research and some enterprise settings, but they have not closed the gap.

Reading notes

#

• Open models sit on a spectrum that runs from open-weight to fully open-source and then to closed API-only models.
• Since about April 2025, Chinese AI companies have been the clear leaders in open-weight models.
• Open-source models that include weights, licenses, inference code, training code, and training data have been led by U.S. groups such as Allen Institute for AI, OpenAthena, and EleutherAI.
• Hugging Face downloads show China taking the lead in July 2025, with the gap widening to about 1.6B downloads.
• Chinese models lead American open models on the Artificial Analysis Intelligence Index, and the top Chinese models are newer than the leading American ones.
• The Chinese open-weight models are estimated to be 2-5 months behind the closed American frontier, while American open-weight models are 6-9 months behind OpenAI and Anthropic.
• Chinese labs release models faster, focus on a narrower task distribution, and also buy more cutting-edge data and RL environments.
• Distillation from American models helps Chinese labs, but the article estimates it only explains a 1-2 month gap.
• Open-weight models are becoming important for cybersecurity risk and raise regulatory uncertainty because software is hard to keep away from bad actors.
• OpenRouter usage grew from about 1T tokens per week to about 80T, and Chinese models rose from about 70% to over 80% of usage.
• Companies and startups such as Harvey, Cursor, DoorDash, Airbnb, and Perplexity are building on Chinese open-weight models.
• Academic research increasingly depends on Alibaba’s Qwen family, with Chinese open models mentioned in over 40% of scanned papers and U.S. models in over 30%.
• The article says the U.S. can still gain ground when its open models are competitive at similar size and capability levels, as seen with gpt-oss, Gemma 4, and Nemotron.
• The closing argument is that open models are becoming the substrate for broader AI diffusion, so the U.S. should keep investing in open models while preparing for their risks.