Meta has lost 11 of the 14 members of the original Llama development team to competitors



Meta will launch '

Llama ' as an open source large-scale language model (LLM) in 2023, and will continue to release new Llama models thereafter, which is at the heart of Meta's AI strategy. However, it has become clear that most of the authors of the papers published when Llama was announced have already left Meta.

Meta Loses Majority of Original Llama AI Team to Competitors - WinBuzzer
https://winbuzzer.com/2025/05/26/meta-loses-majority-of-original-llama-ai-team-to-competitors-xcxwbn/

Meta's Llama AI Brain Drain: Talent Exodus, Model Delays, and the Open-Source Battle Against Rivals
https://www.webpronews.com/metas-llama-ai-brain-drain-talent-exodus-model-delays-and-the-open-source-battle-against-rivals/

exodus at Meta's Llama as 78% of original research team exit
https://www.cryptopolitan.com/exodus-hits-metas-llama-as-researchers-exit/

As Business Insider reports , 11 of the 14 authors of the original Llama paper have already left Meta, raising serious questions about whether the company can remain competitive in the fiercely competitive AI industry, as it struggles to retain top AI talent.

The exodus of AI talent from Meta came after reports that development of Meta's latest model, the Llama 4 Behemoth, was struggling.

Meta plans to release the ultra-large-scale AI model 'Llama 4 Behemoth' with a total of 2 trillion parameters, but development has been delayed until fall 2025 or later - GIGAZINE



Of note, Mistral AI, co-founded by Llama designers Guillaume Lampl and Timothy Lacroix, has poached three of the authors of the Llama paper. Lampl and Lacroix launched Mistral AI in May 2023 and left Meta. Then, in May 2025, they poached Baptiste Rosier, Marie-Anne Lachaud, and Thibaut Lavril, who were also authors of the Llama paper.

In addition, Naman Goyal has joined the Thinking Machines Lab, Aurélien Rodriguez has joined Cohere, Eric Hambro has joined Anthropic, Armand Juhlin has joined Google DeepMind, Gautier Isaaccardo has joined Microsoft AI, and Eduardo Greive has joined Kyotoi .

In addition, Joel Pineau, who has led Meta's basic AI research group for about eight years, will leave Meta in April 2025. In addition, Pineau's successor, Robert Fergus, has worked at Google's DeepMind for five years, 'highlighting the high fluidity of top talent in the field of AI research,' technology media WinBuzzer pointed out.

Regarding Meta's talent exodus, WinBuzzer pointed out, 'This suggests that while Meta continues to invest heavily in AI, its early advantage in the open source space is fading.' 'The departure of most of Llama's original team signals a weakening of Meta's innovation pipeline, which will be even more impactful as many of these researchers join or set up competing AI startups.'



Meta's AI efforts come against the backdrop of a broader corporate restructuring and significant financial commitments: While Meta is aggressively hiring AI and machine learning engineers, it is also cutting staff at its Reality Labs division after posting large operating losses.

Meta lays off dozens of employees of Oculus Studios, its VR/AR division 'Reality Labs,' and hardware business employees may also be subject to job cuts - GIGAZINE



Meanwhile, Meta CEO Mark Zuckerberg stressed the importance of balancing the company's existing platform with future AI development, reportedly saying in an internal meeting that 'we want to reduce the number of abusive posts we detect , while also reducing the number of posts and accounts of innocent people that we mistakenly remove.'

Meta's Llama project is itself navigating a complex landscape: Meta is working on tweaking Llama 4 to address the 'historical leftward shift' it has seen in major LLMs, a trend it attributes to 'the type of training data available on the internet.'

The move comes in tandem with major content moderation policy changes, including the end of third-party fact-checking programs in the U.S. Meta is also facing multiple legal battles, including a copyright infringement lawsuit that claims Llama was trained on a dataset of pirated books.

Meta has trained its AI with 81.7TB of data, including pirated content - GIGAZINE



in Software, Posted by logu_ii