The Current Balance of Power in Open Models

Open-weight, open-source, and closed models

Open language models are AI systems whose weights are publicly available for inspection or downstream use. They are commonly contrasted with closed models, which generally provide access through application programming interfaces (APIs), such as GPT-4 or Claude Opus 4.5, or through products such as ChatGPT and Claude Code.

Open models are usually divided into open-weight and open-source categories. Open-weight models, including Meta’s Llama, Alibaba’s Qwen, Google’s Gemma, and DeepSeek’s models, are currently the more common type. Their licenses define permitted downstream uses, and releases often include inference code through libraries such as Transformers, vLLM, and SGLANG. Since approximately April 2025, Chinese AI companies have led the open-weight model field.

Fully open-source models include the weights, license, and inference code, as well as the information required to reproduce the model, including training code and training data. The most prominent examples have largely come from the United States. These include the Allen Institute for AI’s OLMo models, which were developed during the author’s recent 2.5 years there, along with OpenAthena’s Marin models and EleutherAI’s Pythia models.

These labels describe a spectrum rather than rigid categories. Nvidia’s Nemotron models, for example, are more open than most open-weight models because Nvidia releases substantial amounts of training data under permissive licenses. They are not fully open-source because the complete dataset is not released. Closed models also vary according to the information exposed through their APIs and their terms of use.

U.S.-China competition in open-weight models

GLM-5.2 and Kimi K3, two leading Chinese models, have improved the commercial viability of open models. Their agentic capabilities crossed a threshold similar to the one reached by Anthropic’s Claude Code in December 2025.

The United States initially led open language models, particularly through Meta’s Llama family, which was widely used for research and commercial work. Chinese open-weight models surpassed American models in both research and commercial adoption approximately 18 months ago. Hugging Face download data provides a simple measure of this shift: China took the lead in July 2025, largely because of Alibaba’s Qwen models.

Since the American Truly Open Models (ATOM) Project was first published in August 2025, China’s download advantage has grown to approximately 1.6 billion. Chinese models have received about 3.2 billion downloads in total, twice the American total.

Performance benchmarks show a similar pattern. On the Artificial Analysis Intelligence Index, as of September 14, 2026, the three leading Chinese models were Z.ai’s GLM-5.3 and GLM-5.3-Flash, with scores of 45 and 42, and Moonshot AI’s Kimi K3, with a score of 44. The leading American models were Thinking Machines’ Inkling and Inkling Small, both scoring 26, followed by Nvidia’s Nemotron 3 Ultra, with a score of 23.

The leading American models were released in June and July 2026 and have been updated less frequently than their Chinese counterparts. Chinese laboratories released models that exceeded these American scores two to six months earlier, including GLM-5 and DeepSeek V4 Pro. More American companies, including Arcee AI, Poolside, and IBM, are releasing models, but they are not closing the performance gap quickly. Other benchmarks show a similar trend.

The leading American open models on the Artificial Analysis Index trail 15 other Chinese models. Overall, Chinese open-weight models are approximately two to five months behind the closed American frontier. American open-weight models are approximately six to nine months behind companies such as OpenAI and Anthropic.

The gap is smallest in tasks with clear user demand, including agentic coding, and larger in open-ended scientific work such as physics and biology. Several factors may explain the Chinese advantage. Chinese laboratories often release models more quickly and concentrate on a somewhat narrower range of tasks, which can improve their results on public benchmarks. Because model development progresses continuously, a model released later may represent a stronger snapshot of the field.

These factors do not fully account for the results. Chinese models are technically strong and represent direct competition for American companies. That competition is unlikely to decline substantially as closed laboratories address vulnerabilities in their APIs that allow model distillation.

Distillation is most influential in emerging domains, but it does not make it simple to create a universally capable model. If distillation were fully prevented, for example through know-your-customer (KYC) systems at Anthropic and OpenAI, the gap between the strongest American models and Chinese open-weight models would likely increase by only one to two months.

Chinese laboratories also changed their approach to training data during 2026. Earlier in the year, leading companies such as Moonshot AI and Z.ai generally preferred to build data workflows internally. By summer, they had begun purchasing advanced data, including challenging reinforcement-learning environments for agentic tasks, from established American companies and new Chinese startups.

Safety and regulatory uncertainty

As Chinese open-weight models approach frontier capabilities, concerns are growing about the risks associated with releasing powerful models. Recent incidents have documented risks involving frontier systems and cybersecurity, including the OpenAI-Hugging Face incident.

A structural difficulty is that few effective methods exist for preventing open software from reaching malicious users. Restricting access to the strongest Chinese open-weight models could therefore disadvantage American businesses as well. In one example, Hugging Face used a Chinese open-weight model to analyze a cyberattack because closed models did not answer its requests.

Managing the risks of open-weight models consequently depends in large part on ecosystem-level preparation. Open models are becoming an important mechanism for AI diffusion. Continued investment in American open models could reduce situations in which American companies depend on models developed in China while improving the ability to coordinate responses to risks that cross national borders.

Adoption by companies, researchers, and other countries

Open-weight language models gained substantially in general interest and economic viability during 2026. Platforms such as OpenRouter provide a way to compare usage of American, Chinese, and other models through a single interface. OpenRouter has shared usage data for leading models since January 1, 2025.

Weekly processing by open models grew from approximately 1 trillion tokens in September 2025 to approximately 80 trillion tokens by September 2026. During the same period, Chinese models increased from approximately 70% of usage to more than 80%.

Other platforms built around open models show a similar pattern. OpenCode, an open-source coding agent, reports that approximately 95% or more of its inference volume comes from Chinese models. These platforms offer only an approximation of total usage because services such as Together AI and Fireworks AI do not disclose complete per-model data. Private enterprise deployments also account for a substantial share of usage.

Several prominent technology companies and startups are building AI features on Chinese open-weight models. Examples include Harvey’s legal agent, Cursor’s coding agent, DoorDash’s use of Kimi models, Airbnb’s use of Qwen, and Perplexity’s use of DeepSeek. These companies represent only part of a broader movement among younger Silicon Valley startups seeking lower-cost and more flexible model options.

American companies and startups are also increasingly entering enterprise agreements with Chinese model laboratories to obtain permission to use their models in commercial products. This represents a growing form of cross-border technology collaboration.

Chinese models are also becoming embedded in academic research. Alibaba’s Qwen family is used in much of the research involving open models. To measure this adoption, papers in the five most popular machine-learning categories on arXiv, cs.AI, cs.CL, cs.CV, cs.LG, and stat.ML, were examined.

Mentions of any open model increased from 2% of papers in January 2023 to 50% in September 2026. The leading model families also shifted from the United States toward China.

In April and May 2023, shortly after Meta released the original Llama in February 2023, approximately 2,600 of 12,000 new AI and machine-learning papers mentioned at least one prominent open model family. About 5.5% mentioned Llama, while approximately 1% mentioned a Chinese model.

During Llama’s peak in fall 2024, approximately 23% of papers mentioned Llama and about 7.5% mentioned Qwen. By September 2026, Llama was still mentioned in approximately 21% of papers, showing notable longevity, but Qwen’s share had risen to 30%. More than 40% of papers mentioned at least one Chinese open-weight model, compared with approximately 30% mentioning an American model.

American models with capabilities and parameter sizes comparable to Chinese models are adopted at disproportionately high rates in some cases. OpenAI’s first open-weight models since ChatGPT, gpt-oss, became some of the most adopted open-weight models. Google’s Gemma 4 models have reached adoption levels comparable to Qwen’s most popular small models, while Nvidia’s Nemotron models have achieved modest adoption despite the availability of more capable models of similar size.

Summary of the 2026 landscape

The open-model story in 2026 is increasingly defined by economic relevance.

1. The capability gap between open and closed models has narrowed. The difference varies by task but is estimated at approximately two to five months. Because AI capabilities are advancing quickly, open-weight models are becoming viable for substantial commercial markets.

2. Usage is expanding in high-value industries. Software engineering, legal services, and financial services are among the areas adopting open models. Inference providers such as Together AI, OpenRouter, Fireworks AI, and Baseten are growing alongside a post-training economy that includes fine-tuning APIs such as Thinking Machines’ Tinker. Users often select open-weight models such as GLM-5.3 because of their speed, lower prices, customization options, and privacy characteristics.

3. Chinese AI companies lead the open-weight field. American laboratories have increased their open-model activity since 2025, but Chinese laboratories continue to release stronger models at a faster pace.

4. Distillation does not explain the entire Chinese advantage. Distillation trains one model on outputs from a stronger model. Chinese companies have used techniques to obtain reasoning traces and other data from insufficiently secured American services. Current estimates suggest that distillation reduces the performance gap by approximately one to two months, but it is not the sole explanation for Chinese success.

5. Chinese models, especially Qwen, are becoming foundational research tools. Recent measurements found Chinese open-weight models in 38% of AI papers, compared with 28% for American models, and China’s share is growing faster. This trend, along with political factors and the closed nature of leading American AI companies, is reducing the United States’ lead as the world’s primary AI research hub.

6. Open models are reaching capability levels that introduce new risks. Cybersecurity is one example in which widely available models may enable harmful activity. The resulting risks require ecosystem-level preparation. Researchers and engineers also rely on open models because closed systems such as Claude and GPT may refuse defensive cybersecurity or biology requests.

Conclusion

In 2026, Chinese laboratories clearly lead the open-weight AI ecosystem. This has occurred as open-weight models passed an important threshold in economic viability and as American laboratories increased their own activity. The leading Chinese laboratories continue to expand adoption among enterprises and researchers worldwide.

Open models are becoming a foundation for organizations outside the small group of frontier AI laboratories. They provide a way for businesses and researchers to apply AI to software engineering and other computational tasks. Their distribution also creates influence for the organizations that develop and support them.

The future path of open models remains uncertain. Available data about usage in countries other than the United States and China is limited. Training expertise is distributed across dozens of organizations and thousands of people who are within roughly a year of frontier capabilities. As a result, the question is when, rather than whether, open models will cross performance thresholds that enable new workflows.

The central challenge is to understand how broadly accessible AI can be used constructively while preparing for and reducing its potential harms.