The current balance of power in open models
The expanded form of a testimony I prepared for Congress.
The current landscape of open models in the realm of Artificial Intelligence (AI) is heavily influenced by the ongoing competition between the United States and China. Open models, which refer to AI models with publicly available weights for inspection or downstream use, contrast with closed models that only offer access through APIs or specific products.
Open models are typically categorized into two types: open-weight and open-source models. Open-weight models, such as Meta's Llama, Alibaba's Qwen, Google's Gemma, and DeepSeek's models, are governed by licenses and accompanied by inference code. True "open-source" models, alongside weights and licenses, include the complete information needed to reproduce the model, including training code and data.
China has recently emerged as a leader in open-weight models, with Chinese AI companies dominating this space. The Allen Institute for AI, through its Olmo models, and organizations like OpenAthena and EleutherAI are prominent examples of open-source models in the United States. Open-weight and open-source models represent different levels of openness on a spectrum, with Nvidia's Nemotron models, for instance, being more open than most open-weight models due to the release of large training data sets under permissive licenses. However, they are still not fully open-source as they do not release all the data.
The competition between American and Chinese open-weight models is evident in various aspects, including unit economics, technical capabilities, and capabilities in terms of agentic tasks. Chinese models like GLM-5.2 and Kimi K3 have recently demonstrated significant advancements, matching the capabilities of Anthropic's Claude Code.
In terms of Hugging Face Downloads, a metric tracking usage, China took the lead in July 2025, driven largely by the success of Alibaba's Qwen models. As of September 14, 2026, China leads American models by a substantial margin, with a total of 3.2 billion downloads compared to America's 1.6 billion. Popular benchmarks, such as the Artificial Analysis Intelligence Index (AAII), also show Chinese models such as GLM-5.3, GLM-5.3-Flash, and Kimi K3 leading American counterparts, including Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra.
Despite recent releases by American companies like Arcee AI, Poolside, and IBM, they have not been able to significantly narrow the performance gap with their Chinese counterparts.
Written by urgent.news from Interconnects's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
