Key Takeaways
- The majority of researchers behind Meta’s original Llama AI model have left the company.
- Many have joined or co-founded rival AI ventures, including Paris-based Mistral AI.
- These departures raise questions about Meta’s ability to retain top AI talent and stay ahead in the competitive AI landscape.
- Meta is reportedly delaying a major new AI model, and its latest Llama release received a lukewarm reception.
Meta’s Llama AI models have been central to its artificial intelligence strategy. However, a significant shift has occurred: most of the researchers who developed the groundbreaking original Llama have departed.
Out of the 14 authors credited on the 2023 paper that introduced Llama, only three remain at Meta. The others have moved on, with many now contributing to or establishing competing AI companies.
This “brain drain” is particularly evident at Mistral, a Paris-based startup. It was co-founded by Guillaume Lample and Timothée Lacroix, two key architects of Llama, and now employs several other Meta alumni who are building powerful open-source models.
These exits are creating a buzz about Meta’s capacity to hold onto its leading AI experts. This challenge comes as Meta faces new pressures both internally and from competitors. The Wall Street Journal reported Meta is delaying “Behemoth,” its largest-ever AI model, due to internal concerns about its performance and leadership.
Meanwhile, Llama 4, Meta’s most recent AI release, didn’t quite excite developers as much as anticipated. Many are now turning to faster-moving open-source alternatives like DeepSeek and Qwen for the latest advancements.
Internally, Meta’s AI research team has also seen changes. Joelle Pineau, who led the Fundamental AI Research group (FAIR) for eight years, announced she would step down. Robert Fergus, who co-founded FAIR in 2014 and recently returned to Meta after five years at Google’s DeepMind, will take her place.
The original Llama paper was more than a technical achievement; it legitimized open-source AI models—those with code and parameters freely available for public use and modification—as serious alternatives to closed systems like OpenAI’s GPT-3. Meta even trained its models using only publicly available data.
For a while, it seemed Meta could lead the open-source AI movement. Two years later, that lead has diminished, and Meta is no longer setting the pace. Despite investing billions in AI, the company still lacks a dedicated “reasoning” model, one specifically designed for complex problem-solving or multi-step tasks.
The 11 departed authors had an average tenure of over five years at Meta, indicating they were deeply involved in the company’s AI work. Their departures signal a quiet unraveling of the team that helped build Meta’s reputation in open AI models.
According to Business Insider, these former Meta researchers have found new roles at various AI companies. Many have joined Mistral, while others are now at firms like Google DeepMind, Anthropic, Cohere, Microsoft AI, Kyutai, and Thinking Machines Lab, often in significant positions.
