High turnover hits Meta’s Llama as 78% of the original research team exits

Meta’s pioneering Llama initiative once heralded as a cornerstone of its artificial intelligence roadmap, is now grappling with a significant exodus of key contributors.

Of the fourteen researchers whose names adorn the seminal 2023 paper that unveiled Llama, only three, research scientist Hugo Touvron, research engineer Xavier Martinet, and technical program leader Faisal Azhar, remain at Meta.

The Llama team left for competition

The other eleven team members, or 78% of the researchers, have largely departed to either join or establish rival ventures, leaving Meta’s flagship open-source project without much of its original creative force.

Nowhere is this talent drain more conspicuous than at Mistral, a Paris-based AI upstart founded by Guillaume Lample and Timothée Lacroix, both instrumental architects of Llama’s initial design. Alongside a cadre of fellow Meta alumni, they are hard at work developing new open-source models that directly challenge Meta’s offerings.

This also comes as reports have indicated top AI firms are on a serious talent hunt, paying large sums of money to fish top AI researchers to join their teams.

As for Meta, the migration of expertise has prompted observers to question whether Meta can continue to hold onto top researchers at a moment when the company faces increasing scepticism about its own AI ambitions.

Compounding Meta’s internal challenges, the company recently announced a delay in the release of Behemoth, its largest-ever AI model, in response to concerns raised by employees over its performance and direction.

Meanwhile, developers have remained largely muted at Llama 4, the latest iteration in the model series.

Many are now favoring open-source alternatives, such as DeepSeek and Qwen, that promise more rapid innovation and cutting-edge capabilities.

Could Meta be reorganizing its research department?

The upheaval in personnel has dovetailed with a shake-up in leadership. Joelle Pineau, who for eight years guided Meta’s Fundamental AI Research (FAIR) division, revealed last month that she would be stepping aside.

In her wake comes Robert Fergus, a FAIR co-founder who spent half a decade at Google’s DeepMind before rejoining Meta in May 2025. This transition highlights a broader pattern of turnover and reorganization within Meta’s research ranks.

Since the publication of the Llama paper, FAIR has quietly lost many of its original talents, even as the company continues to spotlight Llama as the lynchpin of its AI strategy.

The question now is whether Meta can defend the lead it once held in open-source model development without much of the team that laid its foundation.

At the time of its release, the Llama paper did more than introduce a new model; it conferred legitimacy upon the concept of openly shared large-language-model weights. Unlike proprietary systems such as OpenAI’s GPT-3 or Google’s PaLM, Llama’s architecture, training code, and parameter sets were freely available to researchers and developers.

Meta demonstrated that by leveraging only publicly accessible data and optimizing for efficiency, state-of-the-art language models could run on a single GPU, democratizing access to advanced AI capabilities.

For a brief period, Meta appeared poised to dominate the open-source frontier. However, two years on, its early advantage has waned. Despite pouring billions into AI research, the company still lacks a dedicated “reasoning” model tailored for tasks that demand multi-step logic, complex problem-solving, or the integration of external tools.

In contrast, competitors such as Google and OpenAI have made these features central to their latest releases, further highlighting Meta’s gap.

The eleven authors who have left Meta each averaged more than five years with the company, indicating a departure of deeply embedded researchers rather than short-term contractors. Their exits span from January 2023 through the Llama 3 cycle and as recently as early 2025, marking a gradual unravelling of the original Llama team.

Meta has acknowledged the departures publicly, with spokespersons pointing to an X post that tracks the career moves of former Llama paper co-authors. While the precise destinations vary, from roles at emerging startups to leadership positions in competing labs, the collective migration reportedly highlights a shift in the AI landscape, where talent follows the most dynamic and open platforms.

Cryptopolitan Academy: Tired of market swings? Learn how DeFi can help you build steady passive income. Register Now