Meta Zuckerberg abandons open source to join Alibaba! Qwen trains Avocado paid AI

Meta CEO Mark Zuckerberg has directed the company to abandon its long-term “fully open-source” AI strategy, with the most advanced AI models in the future adopting a paid model. According to multiple informed sources, Meta’s next-generation large language model (internal code name “Avocado”) is being fine-tuned based on the open-source model series Qwen, developed by Alibaba’s Tongyi Qianwen. Zuckerberg explicitly stated that Meta must prioritize “superintelligence” over a strict open-source stance.

From Leader to Follower: Meta’s Strategic Shift

When Meta launched its flagship Llama series AI models in February 2023, it was regarded by global developers as a leader in open-source artificial intelligence. Zuckerberg publicly claimed in July last year that “Meta is committed to open-source AI, and open-source AI is beneficial to the world.” However, just a few months later, this commitment faced a fundamental change. According to Bloomberg, despite Zuckerberg’s earlier statements about continuing on the open-source path, the Avocado model may be released in a closed-source format.

The core reason for Meta’s strategic shift is the disappointing market performance of Llama 4. Omdia chief analyst Su Lianjie pointed out that Meta’s foundational models underperformed in key industry benchmarks. The Llama 4 model released in April was widely regarded as a failure because it failed to keep pace with the cutting edge of technology. Data from third-party benchmarking firm Artificial Analysis shows that Llama 4’s scores in openness and intelligence fell far short of expectations, contrasting sharply with leading models like OpenAI’s GPT-5 and Google DeepMind’s Gemini 3 Pro.

More critically, Meta found that solely sticking to an open-source strategy could no longer sustain its lead in fierce AI competition. OpenAI achieved rapid commercialization through closed models and continued R&D investment, creating a positive feedback loop. In contrast, although Meta has gained favor with the developer community, it lags behind in technological breakthroughs and commercial realization. Zuckerberg stated in internal meetings that the pursuit of “superintelligence” must take precedence over ideological commitments to open-source.

Alibaba Qwen’s Counterattack

Two years ago, the situation was entirely different. In 2023 and 2024, many Chinese companies eager to catch up with the US heavily relied on Llama to build their own models. At that time, Alibaba’s first-generation Qwen, released in September 2023, adopted Llama’s training process and cited Meta’s pioneering research in its technical reports. Chinese researchers even hailed Llama as a “top-tier open-source large language model,” demonstrating Meta’s absolute leadership in global open-source AI at that time.

However, the situation dramatically reversed in early 2025. In January this year, DeepSeek’s breakthroughs propelled Chinese AI to the forefront of the world, and open-source models like DeepSeek and Qwen began to see widespread adoption globally. According to a US government report in September, in the first nine months of this year, downloads of DeepSeek and Qwen models on the developer platform Hugging Face grew nearly 1000 times and 135 times respectively. Even more astonishing, according to research published last month by Hugging Face and MIT, models made in China accounted for 17% of total open-source model downloads in the past year, compared to only 15.8% for models made in the US.

Three Key Factors That Enabled Qwen to Surpass Meta Llama

Continuous Performance Improvements: Alibaba has continuously enhanced its models over the past year while keeping its CEO Wu Yongming’s promise to keep models open-source.

Balance of Openness and Intelligence: Qwen scores highly on Artificial Analysis’s comprehensive index, balancing openness and performance.

Recognition by Global Developers: Qwen and DeepSeek alone account for over 14% of open-source software downloads in the past year.

Alibaba’s success is no accident. CEO Wu Yongming made a commitment last November to keep its models open-source and continue investing in performance improvements. This persistence has borne fruit this year, with the Qwen model surpassing Llama 4 in key indicators such as multilingual capability, reasoning ability, and code generation. Analysts point out that Meta choosing Alibaba’s Qwen as a base for fine-tuning indicates that Chinese open-source models are now recognized internationally for their performance and multilingual capabilities.

Avocado Project and Meta’s Dilemma

Meta is currently developing the Avocado model, which is being fine-tuned based on Alibaba’s Qwen. Although reports do not specify which version of Qwen is used, the choice itself is highly symbolic. Two years ago, China’s AI industry’s reliance on Llama was seen by experts as a sign that China lagged behind the US in AI. Today, that situation has been completely reversed, with Meta now relying on Chinese open-source models as a foundation.

This strategic shift puts Meta in a difficult position. On one hand, adopting Qwen can accelerate product iteration and commercialization in the short term, leveraging Alibaba’s proven technological path to save significant R&D time and costs. On the other hand, it may undermine Meta’s leadership in the global developer community. For a long time, Meta’s open-source approach has been a core differentiator in its AI strategy. Now, not only has it abandoned open-source in favor of a paid model, but it is also relying on open-source models from its competitors, a shift with implications far beyond technology.

Market reactions quickly reflect this complex sentiment. Following the announcement, Alibaba’s US stock (BABA) rose in pre-market trading, while Meta (META) slightly declined. Investors recognize Alibaba’s technological strength being validated internationally, yet also worry about the uncertainty surrounding Meta’s strategic shift. Analysts believe that while this move may accelerate product development in the short term, in the long run Meta could lose support from the developer ecosystem—a key asset in its competition against OpenAI and Google.