When US tech giant Meta Platforms released its flagship Llama family of artificial intelligence models in February 2023, they were open-sourced, a move that singled it out among global AI model developers at the time.

That September, one of the many derivatives of Llama was announced: Alibaba Cloud’s Qwen. The first generation of Qwen adopted Llama’s training process and cited Meta’s seminal research findings in its accompanying technical report.

Out of deference, the Chinese researchers even called Llama “the top open-source large language model”.

Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.

Two years later, the tables have seemingly turned. According to a Bloomberg report on Wednesday, it was now Meta that is reportedly taking its cues from Alibaba, with unnamed sources claiming that the Facebook owner was now using Qwen to help train a new model code-named Avocado.

The report did not specify which Alibaba Qwen model was being used. Alibaba owns the Post.

Meta released its flagship Llama family of AI models in February 2023. Photo: Shutterstock Images alt=Meta released its flagship Llama family of AI models in February 2023. Photo: Shutterstock Images>

The development caps a remarkable reversal in fortunes for the two leading US and Chinese open-source flagbearers. Up until very recently, Llama was the default model of choice for developers building on open-source AI software, including in China.

Throughout 2023 and 2024, Chinese firms keen to catch up with the US used Llama to help them build their own models. The practice was controversial enough that start-ups like renowned Taiwanese venture capitalist Kai-Fu Lee’s 01.AI received blowback for not adequately revealing the contributions of Llama to their own models.

The Chinese industry’s reliance on Llama at the time was highlighted by experts, including Jeffrey Ding, an assistant professor at George Washington University, as evidence of China lagging the US in AI. Even Chinese industry insiders themselves warned that China’s AI efforts were hindered by its reliance on Meta’s models.

While Alibaba had similarly been open-sourcing its Qwen models since 2023, it always trailed Llama on key performance benchmarks. It took until January this year, when DeepSeek pushed Chinese AI to the forefront, that global adoption of both DeepSeek and Qwen’s open-source models exploded.

According to a US government report in September, downloads of DeepSeek and Qwen models on the developer platform Hugging Face increased nearly 1,000 and 135 per cent, respectively, in the first nine months of the year.

Qwen and DeepSeek alone captured 14 per cent of open-model downloads in the past year. Photo: Shutterstock Images alt=Qwen and DeepSeek alone captured 14 per cent of open-model downloads in the past year. Photo: Shutterstock Images>

The two Hangzhou-based firms have led the way in helping China overtake the US in the global open-source model marketplace for the first time, according to a study published last month by Hugging Face and the Massachusetts Institute of Technology.

Chinese-made models made up 17 per cent of total open model downloads in the past year, the data showed, compared with 15.8 per cent for US-made models. Qwen and DeepSeek alone captured 14 per cent of downloads.

Central to Meta’s demise had been the underwhelming performance of its foundational models on key industry benchmarks, said Su Lian Jye, chief analyst at Omdia. The company’s Llama 4 model release in April was widely seen as a failure as it did not get close to the frontier of the technology.

Meanwhile, third-party benchmark data shows that Alibaba has continuously improved the strength of its models over the past year while staying true to the commitment made in November last year by CEO Eddie Wu Yongming to continue open-sourcing its models.

According to leading third-party benchmarking firm Artificial Analysis, Qwen models scored highly on a new index that ranked models based on their combined openness and intelligence, in contrast to leading closed US models such as OpenAI’s GPT-5 and Google DeepMind’s Gemini 3 Pro, which had top scores for intelligence but low scores for openness.

In contrast, Meta was now looking set to abandon its open-source strategy, with Bloomberg reporting that the Avocado model could be released as a closed model despite CEO Mark Zuckerberg claiming in July last year that his company would continue on the open-source path.

“Meta is committed to open source AI,” he wrote at the time. “Open source AI is good for the world.”

This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice reporting on China and Asia for more than a century. For more SCMP stories, please explore the SCMP app or visit the SCMP’s Facebook and Twitter pages. Copyright © 2025 South China Morning Post Publishers Ltd. All rights reserved.

Copyright (c) 2025. South China Morning Post Publishers Ltd. All rights reserved.