The Shifting Sands of AI: From Chinese Adoption of Llama to Meta’s Embrace of Qwen
The global artificial intelligence landscape is in constant flux, with power dynamics and technological leadership evolving at an unprecedented pace. For a significant period, Meta Platforms’ Llama family of AI models served as a cornerstone for innovation, particularly within the burgeoning Chinese AI sector. However, recent developments suggest a remarkable turn of events, with Meta now reportedly drawing inspiration from Alibaba’s Qwen models, reversing the earlier trend of Chinese firms leveraging Meta’s open-source technology.
The Rise of Llama and Early Chinese Adoption
When Meta first unveiled its Llama AI models in February 2023, their open-source nature set them apart in a field dominated by proprietary systems. This accessibility quickly made Llama a favored tool for developers worldwide, including those in China. By September of the same year, Alibaba Cloud announced its Qwen model, a derivative that openly acknowledged its reliance on Llama’s training methodologies and cited Meta’s foundational research. The respect for Llama was so profound that Chinese researchers referred to it as “the top open-source large language model.”
During 2023 and into 2024, Chinese companies actively utilized Llama to accelerate their own AI development efforts. This reliance was so pronounced that it sparked debate, with some startups facing criticism for not fully disclosing Llama’s contribution to their proprietary models. Experts, such as Jeffrey Ding, an assistant professor at George Washington University, pointed to this dependence on Meta’s models as an indicator of China’s then-lagging position in the AI race compared to the United States. Even within China’s tech industry, concerns were raised that this reliance might be hindering independent innovation.
The Emergence of Qwen and DeepSeek
While Alibaba had been releasing its Qwen models under open-source licenses since 2023, they consistently lagged behind Llama in key performance metrics. The landscape began to shift significantly in January of this year. The breakthrough came with the increased global adoption of open-source models from both DeepSeek and Qwen, propelling Chinese AI capabilities to the forefront.
A report from the US government in September highlighted this surge in popularity, indicating that downloads of DeepSeek and Qwen models on the developer platform Hugging Face had seen dramatic increases of nearly 1,000% and 135%, respectively, within the first nine months of the year.
China Takes the Lead in Open-Source Models
This impressive growth has positioned China as a leader in the global open-source model marketplace, a significant milestone achieved for the first time. A study released last month by Hugging Face and the Massachusetts Institute of Technology revealed that Chinese-made models accounted for 17% of all open model downloads over the past year, surpassing the 15.8% share held by US-made models. Qwen and DeepSeek alone contributed a substantial 14% of these downloads.
Meta’s Potential Pivot and the Future of Open Source
Several factors are believed to have contributed to Meta’s perceived decline in this arena. According to Su Lian Jye, chief analyst at Omdia, Meta’s foundational models have underperformed on critical industry benchmarks. The release of Llama 4 in April was widely considered a disappointment, failing to reach the cutting edge of AI technology.
In contrast, Alibaba has demonstrated a consistent improvement in its models’ capabilities throughout the past year. This progress aligns with the company’s commitment, reiterated by CEO Eddie Wu Yongming in November last year, to continue its open-sourcing strategy.
Third-party benchmarking firm Artificial Analysis has recognized the strength of Qwen models. They have scored highly on a new index that evaluates models based on a combination of their openness and intelligence. This stands in stark contrast to leading closed-source US models like OpenAI’s GPT-5 and Google DeepMind’s Gemini 3 Pro, which excel in intelligence but score poorly on openness.
Intriguingly, Meta appears to be considering a departure from its open-source approach. Bloomberg reports suggest that the company’s new model, codenamed Avocado, might be released as a closed model. This potential shift contradicts the stance taken by CEO Mark Zuckerberg in July last year, when he emphasized Meta’s commitment to open-source AI, stating, “Open source AI is good for the world.” This potential move by Meta could signal a significant change in its strategy and impact the broader open-source AI ecosystem.
