The narrative in Western AI circles is that the United States is winning the open source AI race. Meta has Llama. Google has Gemma. Microsoft has Phi. The occasional Mistral release from France rounds out the story. But the 2025 numbers tell a different story. Chinese companies and research institutions released 89 significant open weight AI models (models whose trained parameters are publicly downloadable, as opposed to closed weight models that can only be used through an API) in 2025, compared to 47 from American organizations. The same gap held in the first quarter of 2026.
The gap is not just about the raw count. It is about the kinds of models being released, the licensing terms, and the research community’s access to the training data and the training recipes. On all three axes, Chinese open source AI is ahead of American open source AI right now. The reasons are mostly about state policy and the commercial realities of the Chinese AI market, not about technical superiority.
I have downloaded and tested most of the major Chinese open weight models from 2025 and Q1 2026. I have written code against the DeepSeek R1 reasoning model, the Qwen 3 multimodal model, the Yi-1.5 large language model, and the Hunyuan translation model. I have also shipped production code against the Llama 3.x and Gemma 3 line from the US side, and against Mistral and Mixtral from Europe. The gap is real. The reasons are not what the Western press has been writing about.
What got released
The Chinese open weight model releases of 2025 that mattered:
- DeepSeek R1 in January 2025. A reasoning model (an AI trained to produce step-by-step intermediate reasoning before its final answer, similar in concept to OpenAI’s o1 series) that matched OpenAI’s o1 on the standard reasoning benchmarks while being open weight and free to download. The model card (the documentation file that ships with a public model describing what it was trained on and how it should be used) was the first one I had seen from a Chinese lab that documented the training data composition at the level of detail that American labs document. DeepSeek also published the training recipe (the step-by-step procedure used to train the model, including data sources and hyperparameters) and the reinforcement learning from human feedback (RLHF) code, which is a level of openness that Llama 3 does not match.
- Qwen 3 in April 2025. A multimodal model (an AI that can process more than one type of input, such as both text and images) from Alibaba that handles text, images, and audio in a single model, with open weights under the Apache 2.0 license (a permissive open-source license that allows commercial use with no copyleft restrictions). The model is smaller than the equivalent Llama 3 multimodal model and benchmarks competitively on most tasks I tested.
- Yi-1.5 34B and 9B from 01.AI, also April 2025. The 34 billion parameter model is the strongest Chinese model for code generation in the small-to-mid size class. The 9 billion parameter model is the one I have been using on my laptop for offline coding assistance.
- Hunyuan translation model, June 2025. A specialized translation model from Tencent that outperforms the general-purpose models on the standard translation benchmarks for Chinese to English and English to Chinese, and is open weight.
- DeepSeek V3 in December 2025. A 671 billion parameter mixture of experts model (an architecture where only a small subset of the model’s “experts” or sub-networks are activated for any given input, making large models cheaper to run) that matches the closed weight frontier models on the major reasoning benchmarks. The training cost disclosure (around $5.5 million, compared to estimates of $100 million+ for the equivalent American closed weight models) was the part that broke the Western press cycle for two weeks.
The American open weight releases of 2025 that mattered in comparison:
- Llama 3.3 70B in December 2024, the last major Llama release of 2025. Strong model, closed training data, restrictive license (the Llama community license has commercial restrictions that the Apache 2.0 license does not).
- Gemma 3 in March 2025. Google’s open weight line, smaller models, strong on the benchmarks, restrictive license (the Gemma license has red lines around the size of derivatives you can ship).
- Phi-4 in December 2025. Microsoft’s small language model line, around 14 billion parameters, very strong for the size. Closed training data, MIT license (one of the most permissive open-source licenses, allowing nearly unrestricted commercial use).
- Mistral Small 3 in November 2025. Apache 2.0 license, around 22 billion parameters, weaker on the reasoning benchmarks than the Chinese equivalents at the same size.
Counting all of the above plus a handful of smaller releases from Allen AI, Hugging Face, and academic labs, the American open weight count for 2025 is 47. The Chinese count is 89 if you include the major releases from DeepSeek, Alibaba, 01.AI, Tencent, Baidu, the Chinese Academy of Sciences, and a long tail of university and startup releases.
Why the gap is real and not a counting trick
The first objection I get from people in the Western AI community is that the count includes some models that are not really “open.” The Chinese models ship the weights (the numerical parameters that define a trained neural network) but some of them do not ship the training code or the training data. This is a fair point. The Open Source Initiative’s definition of “open source AI” (the standard most Western open source advocates use) requires the training data and the training code, not just the weights. By that strict definition, the Chinese count drops from 89 to about 60, and the American count drops from 47 to about 30.
But even with the stricter count, the Chinese releases are larger and they are more open about the training data composition. The DeepSeek R1 model card documents the training data sources, the data filtering steps, and the human evaluation protocol in a way that the Llama 3 model card does not. The Qwen 3 release includes the data preprocessing code. The Yi-1.5 release includes the benchmark evaluation harness. None of the American open weight releases include all three.
The Chinese open weight models also tend to ship with less restrictive licenses. Apache 2.0 and MIT (both permissive open-source licenses with minimal restrictions on commercial use) dominate the Chinese release set. The American open weight set is split between Apache 2.0 and the more restrictive Llama community license and the Gemma license. For a company that wants to ship a product built on top of an open weight model, the Chinese releases are the easier starting point.
Why the gap exists
The reasons are mostly about state policy and market structure, not about technical superiority.
State policy in China has identified open source AI as a strategic priority. The 2024 work recommendations from the Ministry of Industry and Information Technology explicitly called for open source AI development as part of the broader AI strategy. State funding for open source AI research has been consistent since 2022, and the funding goes to both the big tech companies (Alibaba, Tencent, Baidu, DeepSeek) and to the academic research labs (the Chinese Academy of Sciences, Tsinghua, Peking, a long tail of second-tier universities).
American state policy has taken the opposite approach. The 2023 executive order on AI (an early Biden-administration directive that shaped federal AI policy) and the subsequent 2024 National Security Memorandum explicitly restricted some categories of open weight AI release on national security grounds. The CHIPS and Science Act (a 2022 US law that funded domestic semiconductor research and restricted advanced chip exports to China) and the subsequent export controls on advanced GPUs to China were justified on national security grounds, but the same policy framework has made American open weight releases more cautious. The Llama 3 license, for example, explicitly bans use by companies above a certain user count threshold in some sectors, which is a direct response to the regulatory pressure.
Market structure is also different. The Chinese AI market is dominated by companies that have to compete with each other on open source releases because the closed weight frontier model market is dominated by American companies (OpenAI, Anthropic, Google) that Chinese companies cannot easily compete with. For Alibaba and Tencent, the open weight release is a way to build developer mind share (the degree to which software developers are aware of and prefer a given tool) and to recruit researchers. For Meta and Google, the open weight release is a defensive move to keep developers from going to the closed weight API providers, and the defensive motivation is weaker because the closed weight providers are also American.
The net effect is that the Chinese open source AI community is more incentivized to release, and the American open source AI community is more incentivized to hold back. The result is the gap we see in 2025 and Q1 2026.
What it means for someone shipping AI products
If you are shipping a product that uses an open weight model as the inference engine (the component that actually runs the model to generate predictions or text), the practical question is which model to standardize on. The Chinese releases have the better licenses, the better documentation, and the better benchmark performance at the same size. The American releases have the better community in the West, the better integration with the American AI tooling, and the lower geopolitical risk (the concern that a model trained in China might be subject to a Chinese government request for the training data, the model weights, or the user data).
I have been standardizing on Qwen 3 for the multilingual production workloads, Yi-1.5 9B for the local-only offline coding assistance, and DeepSeek R1 for the reasoning-heavy batch jobs. The closed weight API providers (OpenAI, Anthropic, Google) are still in the mix for the workloads where the cost of the API is justified by the quality, but the open weight stack is the default for everything else. The American open weight models (Llama 3.3, Gemma 3) are still in the rotation for the workloads where the American tooling integration is a hard requirement.
The geopolitics are real. The US government has been signaling that it may restrict the use of Chinese open weight models by US companies, similar to the TikTok restrictions. As of June 2026, no such restriction is in place, but the signal is there. If you are a US company standardizing on a Chinese open weight model, you should be tracking the policy debate. If you are a non-US company, the policy debate is less of a concern and the model choice is mostly about the technical and licensing trade-offs.
Trade-offs
The trade-offs in standardizing on a Chinese open weight model are:
- Geopolitical risk. A US company standardizing on a Chinese model is exposed to the policy risk. A non-US company is exposed to the same risk if it has US customers, US investors, or US infrastructure.
- Community and tooling. The American open weight models have a larger Western developer community, more Western tooling integration, and more Western language documentation. The Chinese open weight models have a larger Chinese developer community and better Chinese language documentation, but the English language documentation has been catching up fast.
- Support contracts. Neither the Chinese nor the American open weight releases come with a support contract. The closed weight API providers (OpenAI, Anthropic, Google) do. For workloads where the support contract matters, the closed weight API is still the right choice.
The trade-off I would not accept is the one where you standardize on a Chinese model because it is free, without thinking through the geopolitical risk or the community gap. The cost savings are real, but they are not the only consideration.
Bottom line
Chinese open source AI is ahead of American open source AI on the count of releases, on the openness of the training data documentation, and on the restrictiveness of the licenses. The gap is a result of state policy and market structure, not technical superiority. If you are shipping AI products and you are not tracking the Chinese open weight releases, you are missing the strongest models in the open weight category right now.
If you only do one thing from this article, download the Qwen 3 32B model and run it locally on a workstation with a 24GB GPU. The license is Apache 2.0, the model is multimodal, the documentation is good, and the benchmark performance is competitive with the closed weight API providers at a fraction of the cost. The geopolitics are a real concern, but the technical reality is that the Chinese open weight models are the strongest open weight models available in 2026.