Silicon Valley startups are turning to Chinese open-source AI models

Misha Laskin, a theoretical physicist and machine learning engineer who contributed to the development of some of Google’s most powerful AI models, encountered a concerning picture when examining the American AI landscape earlier this year.

Laskin observed a growing interest among US AI companies in free, customizable, and increasingly powerful open-source AI models.

The vast majority of these models are produced in China and are rapidly gaining ground against their US competitors.

Assessing the current situation, Laskin stated, “These models are not far behind the frontier (the cutting edge of technology). In fact, they are surprisingly close to the frontier. What is coming now is noticeably close to the frontier.”

Following this development, Laskin founded a startup called Reflection AI to offer an open-source American alternative to the Chinese models gaining traction in Silicon Valley.

The founder of the company, which recently reached an $8 billion valuation, said, “You are starting to see signs that open model companies in China are actually pushing the frontier of intelligence and the limits of intelligence technology in general.”

Over the past year, a significant portion of America’s most popular AI startups have turned to Chinese open AI models, which compete with and sometimes replace expensive US systems as the foundation for American AI products.

More than 15 AI startup founders, engineers, and industry experts who spoke to NBC News stated that American companies’ models still hold the lead in terms of capability.

However, experts emphasized that many Chinese systems are cheaper to access, more customizable, and have become sufficiently competent for many use cases over the past year.

Cost and speed advantages are changing preferences

Investors have poured tens of billions of dollars into OpenAI and Anthropic with the expectation that leading American AI companies will dominate the global market.

But the increasing use of free Chinese models by American companies raises questions about how exceptional these models are and whether America’s insistence on a “closed model” approach is flawed.

Michael Fine, head of machine learning at the search company Exa, which is valued at $700 million and backed by established Silicon Valley investors Lightspeed Venture Partners and Nvidia, said that running Chinese models on their own hardware is, in many cases, much faster and cheaper than using large models like OpenAI’s GPT-5 or Google’s Gemini.

Fine described the process:

“We often launch a feature with a closed model, but then we realize it’s too expensive or too slow, and we ask, ‘What tricks do we have up our sleeve to make this faster and cheaper?'”

Fine stated that the solution is often to replace the closed model with an equivalent open model and then run it on their own infrastructure.

Chinese-origin systems like DeepSeek’s R1 and Alibaba’s Qwen models can be used for free because they are “open-source” or “open-weight,” meaning anyone can download, copy, modify, and run them.

These systems differ from “closed” systems accessed through data centers controlled by major tech giants, such as Anthropic’s Claude or OpenAI’s GPT models.

The technology gap is closing fast

For years, the closed-source models from OpenAI and Anthropic performed far better than both American and Chinese open alternatives.

Even open-source initiatives like BloombergGPT, trained by institutions with resources like Bloomberg on their own financial data, lagged behind OpenAI’s closed models in financial knowledge.

However, over the past year, Chinese companies like DeepSeek and Alibaba have made significant technological strides. According to metrics tracked by Artificial Analysis, an independent AI benchmarking company, their open-source products now approach or match the performance of leading closed American models in many areas.

“The gap is really narrowing,” said Lin Qiao, co-creator of PyTorch, the dominant framework for training AI models, and CEO of Fireworks AI, regarding the capability difference between American closed-source and Chinese open-source models.

As a result of this performance increase, platforms like OpenRouter, which allow users to choose between different models, are seeing a shift toward Chinese open-source models.

Jerry Liu, founder of the productivity app Dayflow, estimates that about 40% of his users now prefer to use open-source models.

Dayflow offers an application built on basic tasks like scanning screenshots and summarizing user activity.

Users can choose between Google’s Gemini model and smaller open-source options like Alibaba’s Qwen.

Liu noted that for tasks like describing a user’s screen, the Qwen model is extremely consistent, stating, “Qwen is as good as GPT-5 for my use case.”

Unlike GPT-5 or Gemini, a smaller version of Qwen can be run at a relatively low cost or for free.

Liu mentioned that paying for closed model usage could cost Dayflow up to $1000 per person, making cheaper open-source models critical for the application’s sustainability.

Privacy sensitivity encourages local processing

The open-source models used by Dayflow perform all processing on each user’s own computer. Liu stated that this is attractive to users who do not want to send their data to the cloud for privacy reasons.

Emphasizing his preference for using open-source models on his own device, Liu said, “Would I use a product where my entire screen is beamed to some random guy’s cloud? Never.”

In addition to increased performance, stronger privacy, and lower costs, open-source models are also gaining ground due to ecosystem advantages.

The rising adoption rate among developers and the open-source systems they create encourage more software engineers to use these models.

Antonio Vespoli, co-founder of the browser assistant startup Circlemind AI, said that Chinese models now dominate online developer resources.

There is a practical reason for this: Chinese models like Qwen, which Airbnb CEO Brian Chesky stated they rely on “heavily,” have abundant training guides and community support.

Charles Zedlewski, chief product officer at the AI infrastructure company Together AI, noted that developers now find it simpler and more efficient to start with open models and adapt them with their own data.

Zedlewski stated that companies understand their needs more clearly as they launch their first AI applications.

Of the top 20 models among users of Kilo Code, a popular application that helps software engineers write code, seven are of Chinese origin, and six of them are open-source.

Beijing’s strategic support and production speed

While most of America’s AI developments occur in the private sector and with a closed-model approach led by industry giants like OpenAI and Anthropic, the Chinese government plays a more active role in charting the country’s AI vision.

In a speech on November 1, Chinese President Xi Jinping called for “more cooperation in open-source technologies.”

In March, China’s top economic planning authority announced its intention to support an ecosystem of open-source models.

While Chinese labs generally release their models openly, American companies like OpenAI achieved early success with closed models and have remained committed to that approach.

Furthermore, many Chinese companies are releasing their products at a faster pace than their American competitors.

Alibaba has released a new model roughly every 20 days this year, while the average time between Anthropic’s releases has been 47 days.

Nathan Lambert, a senior research scientist at the Allen Institute for AI and an expert on the open model ecosystem, told NBC News that the recent progress of Chinese models is no coincidence.

“The Chinese are real innovators in AI,” Lambert said.

Lambert, who writes extensively about China’s AI developments on the Substack platform and is considered an expert on China’s open-source ecosystem, added that the balance of power has shifted rapidly in the last 12 months.

Some in Silicon Valley note that American models still hold a significant advantage at the cutting edge of AI capabilities and that closed American models offer a user-friendliness that cumbersome open models cannot match.

Tim Tully, a partner at Menlo Ventures, argued that closed models are still much more capable and generally more useful:

“The tools are better, the productivity is better, the agent frameworks being built and used by everyone are better with Anthropic and OpenAI. They just work better. So the ecosystem is strong in the closed-source environment.”

However, many companies may avoid using Chinese models due to the real or perceived risks of using a product built on a Chinese-origin foundation.

“There is a perceived risk that buyers, whether from the private or public sector, are hesitant to purchase a product based on a Chinese-origin open-weight model,” said Tully, an investor in Anthropic, one of the world’s leading closed-model companies.

The US open-source ecosystem is waking up

American AI companies and the federal government have taken notice of the recent rise of Chinese models. Experts have described America’s lack of powerful open-source models as an “existential” threat to democracy.

Although Meta’s high-profile Llama series has historically led American open-source efforts, CEO Mark Zuckerberg has signaled that Meta does not intend to make all of its “superintelligence” AI models open-source.

The stagnation in the performance of Llama models in recent years is also seen as one of the reasons open-source users have shifted to better-performing Chinese models.

But the US open-source ecosystem may be gradually awakening, with efforts by American innovators to enhance their competitiveness.

In July, the White House released an AI Action Plan that called on the federal government to “Promote Open-Source and Open-Weight AI.”

In August, OpenAI, the creator of ChatGPT, released its first open-source model in five years. Announcing the model’s launch, OpenAI referenced the importance of American open-source models, stating, “Broad access to these capable open-weight models created in the US helps expand democratic AI.”

The Seattle-based Allen Institute also released its latest open-source model, Olmo 3, at the end of November, designed to help users “quickly build reliable features for research, education, or applications,” according to the launch announcement.

Lambert from the Allen Institute also launched the “ATOM Project” (American Truly Open Models).

The ATOM Project’s manifesto states: “America has lost its lead in both performance and adoption in open models and is on track to fall further behind.”

“If we want to be the leading nation in the age of AI, we cannot cede such a critical piece of the ecosystem to any one nation,” Lambert said in a statement to NBC News.