Reversing the celestial order, Meta copied Alibaba's Tongyi Qianwen without authorization.

On December 10th, it was reported that Meta incorporated Alibaba's Tongyi Qianwen model into the development of its new-generation large model, "Avocado," to fine-tune and optimize the new model.

Tech Daily Push exclusively learned from Alibaba Cloud that Meta did not seek authorization from Alibaba in advance, and Alibaba only found out last night.

It is reported that Alibaba's Qianwen is an open-source model, and anyone can download it independently without authorization. According to official disclosures, as of now, the cumulative global downloads of the Qianwen model have exceeded 700 million.

Meta Incorporates Alibaba's Qianwen Model, Once the King of Open Source

Since this year, after the underwhelming performance of Llama 4, the gap between Meta and its competitors like OpenAI and Google has gradually widened. Against this backdrop, the Avocado project emerged.

As Meta's next-generation flagship large AI model, the "Avocado" large model is regarded as Meta's "saving grace" in the AI arms race, with its target performance aimed directly at GPT-5. It is planned to be released in the first quarter of 2026.

At that time, overseas netizens made a meme: the core of the avocado is DeepSeek.

However, when the "avocado" was finally "cut open," it turned out to be Qianwen instead of DeepSeek.

It should be noted that in the past two years, Meta's Llama model dominated the global open-source community. At that time, many large domestic models were often suspected of being rebranded Llama models. Who would have thought that a few years later, Chinese open-source models would rise, and Meta would end up being the one actually "rebranding."

Moreover, Meta has completely deviated from its original intention and abandoned the open-source path.

The Avocado large model will shift from the open-source path of the Llama series to a closed-source proprietary model, only accessible to enterprise users through API interfaces and hosting services.

It uses the open-source Qianwen model for training but then goes closed-source to charge fees. Although it doesn't violate any agreements legally, it's a bit morally questionable.

The model is being developed by a team led by Meta's new Chief AI Officer, Alexander Wang, with Meta founder and CEO Mark Zuckerberg personally overseeing the project.

Two Leaders of Meta AI, One on the Fringes, One Inexperienced

When discussing Meta's AI strategy, two key figures are unavoidable - Yann LeCun and Alexander Wang.

Among them, Yann LeCun, the founder of Meta AI, announced his departure from Meta last month.

In 2013, Yann LeCun joined Facebook, the predecessor of Meta, and founded the Artificial Intelligence Research Institute. He later served as Meta's Chief AI Scientist, leading all of Meta's AI research directions and spearheading the development of the Llama series of open-source large models. He won the Turing Award in 2018.

Although he is undoubtedly a titan in the field, Yann LeCun's AI concepts fundamentally differ from the mainstream views in the industry.

While the industry generally believes that large language models are the core path to AGI, he is skeptical, arguing that although existing large language models perform well in tasks such as language processing and exam problem-solving, they lack common-sense understanding of the physical world and cannot learn physical common sense from limited experiences like animals, leading to "hallucinations" during the generation process.

Therefore, he proposed that the core architecture of the next generation of AI should be a "world model" that understands the world through multi-modal perception (vision, touch, hearing, etc.).

Looking back now, with the emergence of video generation models, more and more researchers are starting to recognize the importance of understanding the physical world. However, his concept was quite forward-thinking at the time.

Sometimes, being too far ahead can cause one to miss out on development opportunities. For example, Meta didn't benefit much from this year's AI boom but spent a lot of money, such as poaching talent at high salaries.

Alexander Wang is the AI talent that Zuckerberg poached at a high price. In June 2025, Meta acquired a 49% stake in Alexander Wang's AI data annotation company for $14.8 billion (approximately 104.5 billion yuan).

In July, he was appointed as the Chief AI Officer of Meta's Super Intelligence Laboratory. At only 28 years old, he leads Meta's closed-source AI development and has real power. Even Yann LeCun's lab reports to him.

It's quite a sight to think that a 65-year-old Turing Award winner and industry veteran has to report to a 28-year-old college dropout.

Although Alexander Wang is indeed talented and became a billionaire at a young age, his experience pales in comparison to other top AI leaders in Silicon Valley.

Demis Hassabis, the founder and CEO of Google DeepMind, is the father of AlphaGo and a versatile genius. He won the Nobel Prize in Chemistry in 2024 for his contributions to "protein structure prediction."

Both OpenAI and xAI are directly led by their founders in AI projects. Compared to industry titans like Sam Altman and Elon Musk, Alexander Wang seems a bit out of his league.

Meta seems to have realized this and has been aggressively poaching talent from companies like OpenAI at high salaries. So far, it's unclear how much this will improve R & D, but many people have achieved financial freedom through this.

Chinese AI Goes Global, A Bipolar World in Ten Years

Meanwhile, Chinese self-developed AI models are expanding globally.

On November 25th, Singapore's National Artificial Intelligence Program announced that it would abandon Meta's Llama series of models in its latest Southeast Asian language large model project and fully switch to Alibaba's Qianwen open-source architecture.

Earlier, Airbnb CEO Brian Chesky said in a media interview that the company relies heavily on Alibaba's Qianwen model, stating that it is "better and cheaper than OpenAI."

Meta went from the top to the bottom in just two or three years. Imagine what the global AI landscape will look like in ten years with the rapid development of Chinese AI companies?

Boldly predicting, considering cost-effectiveness, more developing countries in Africa, Southeast Asia, and Latin America may adopt Chinese AI technology. European countries that often align with the US and wealthy Middle Eastern countries may choose US AI technology due to concerns about data privacy and ethical standards.

By then, China's market share in the global AI market will increase from the current 30% to 40% - 45%, while the US will decline from about 50% to 45% - 50%, forming a true "bipolar world" in the global AI landscape.

In terms of global market share, OpenAI and Google may take the top two spots in the US, accounting for about 30% of the market. In China, companies like Alibaba, ByteDance, and DeepSeek may lead the way, together accounting for 25% of the market.

However, given the rapid pace of AI development, ten years is a long time, and a new giant may emerge at any time. Regardless, it's clear that China is gradually gaining the upper hand.

This article is from the WeChat official account "Tech Daily Push" (ID: apptoday), written by Zhao Zhishan, and is published by 36Kr with authorization.