HONG KONG, CHINA - JANUARY 28: In this photo illustration, the DeepSeek app is seen on a phone in front of a flag of China on Jan. 28, 2025, in Hong Kong, China. (Photo illustration by Anthony Kwan/Getty Images)
Getty Images
When two mysterious and powerful open source AI models appeared on OpenRouter earlier this month without developer attribution, users quickly began speculating on who was behind the release. At the time, there were rumors that this could be a new release from DeepSeek.
While Xiaomi has since come out and claimed the models as Mimo V2 Pro and Mimo V2 Omni, the episode demonstrates the growing hype behind Chinese open source AI models, which were on fire throughout 2025. One of the highlights included DeepSeek R1 going viral after offering comparable performance to OpenAI’s o1.
The momentum of Chinese open source AI accelerated throughout the year. According to a report released by Hugging Face, a platform for sharing open weight AI models, Chinese open source models surpassed the U.S. in both monthly and overall downloads. The majority of trending models in 2025 were either developed in China or a derivative of a model developed in China.
But why are these models doing so well in the market? There are many reasons, but one of the most significant is that they offer robust performance at a fraction of the cost of leading proprietary models like ChatGPT and Claude.
At the time of writing in March 2026, a modified and fine-tuned version of Alibaba’s Qwen sits at the top spot on Hugging Face’s open LLM leaderboard. Across the market, we’re starting to see Chinese models like DeepSeek, Kimi and GLM-5 gain significant momentum.
“Right throughout 2025 and beginning of 2026 their adoption has increased tremendously,” said Clément Delangue, cofounder and CEO of Hugging Face, who spoke to me in a video interview to give his thoughts on what’s behind the momentum of Chinese open source models. “For the first time in 2025, the volume of downloads of models from Chinese model providers has surpassed the numbers of downloads from American model providers.
“I think they’re investing a lot in the topic, in AI in general, and they’re sharing much more openly than in the U.S., right? I think for Chinese AI labs right now, the default is open source. The standard is open source versus in the U.S., the default and the standard is closed source,” Delangue added. “The volume of artifacts, of data sets, of models that are shared by Chinese players right now is just much higher than what’s happening in the U.S.”
It’s not just Hugging Face’s data that points to the dominance of these models either. OpenRouter, a unified API platform with over 5 million users, maintains an LLM leaderboard and the top five ranking models: Mimo V2 Pro, Step 3.5 Flash, Deepseek, V3.2, MiniMax M2.5 and GLM 5 Turbo are all made by Chinese companies.
“Open source models have a few advantages regardless of their origin. The cost is significantly lower. Hosting your own models or using MoE models can reduce cost by 5-10x,” Steve Frey, cofounder and product lead at AGI, Inc. told me via email. “What makes these models useful is customization. Open weights allow models to be adapted and tweaked for specific uses, which is why Qwen has over 113,000 variations on Hugging Face.”
As more developers experiment with these models, we’re starting to see companies like AirBnB, Pinterest and Notion incorporate Chinese models into the tech stack. For instance, AirBnB’s cofounder and CEO Brian Chesky has confirmed the company uses Qwen to power its AI customer service chatbot.
Pinterest is one of the biggest websites in the world, with 619 million monthly active users sharing images across its visual discovery engine. In an email interview with me, Matt Madrigal, CTO of Pinterest, confirmed that the company has been experimenting with a mix of in-house and open source models across its recommendation and AI systems.
“We use open source models where they add value, and then adapt or combine them to improve overall performance and efficiency. In fact, our proprietary multimodal AI model is trained using open-weight model optimization techniques, resulting in it outperforming off-the-shelf models by 30% on shopping relevancy,” Madrigal said.
Madrigal singled out Qwen as an example of an open-source model that works well alongside the company’s in-house models. More specifically, helping to improve content understanding, handling complex multimodal queries and experimenting with new search and assistant experiences.
Pinterest Assistant, the company’s conversational AI assistant, also combines internally built retrieval, recommendation and generative systems with multimodal visual search that help interpret queries, plan responses and call the right tools. Similarly, an internal framework called Navigator-1 combines visual embeddings from the company’s taste graph alongside fine-tuned open source models like Qwen to power AI-driven experiences.
“We started by evaluating both third-party proprietary and open source models. What we found was that recent advances in open source had significantly narrowed the performance gap with leading closed models. From there, it made sense for us to invest more in open-source models and in custom in-house models tailored specifically to Pinterest’s specific needs,” Madrigal said.
In my interview with Delangue, he also briefly mentioned that Hugging Face had been experimenting with open source models internally, with a mix of American, Chinese and French models. He notes that the LLM hosting provider uses Chinese open source models to provide automatic translation of community content posted on the platform, to translate from English to Mandarin, although he didn’t specify which model the company uses.
While Chinese open source models are advancing rapidly, U.S. closed source models remain at the top of the industry, with ChatGPT at 900 million weekly active users and Google Gemini at 750 million monthly active users.
A research paper published by MIT Sloan looked at data on model-level usage on OpenRouter and also found closed source models were dominating the market, accounting for roughly 80% of model usage, even though they cost six times as much as open source AI models.
S0 for now, closed source AI, particularly in the U.S., remains strong. “The gap between open and closed will keep narrowing but never fully close. Proprietary labs will always invest in alignment, safety research and capability work that doesn’t need to be immediately open-sourced. But open models will be sufficient for many everyday tasks,” Frey said.
There are also concerns about the data privacy of Chinese open source models. In one notable example, South Korea alleged that DeepSeek transferred user data to China without user consent, which resulted in the app being removed from the country’s app stores. Likewise, legal requirements forcing Chinese companies to share data with the PRC also raise questions about how secure apps built with these open source models really are.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。