Meta’s once-promising Llama AI initiative faces a significant challenge as the team behind its development experiences a substantial exodus of talent.

According to Business Insider, 11 out of the 14 creators of Meta’s AI model Llama have departed the company, with French AI startup Mistral successfully recruiting five of these former team members.

This talent drain comes at a critical juncture for Meta, which has been aggressively pushing its AI capabilities through the Llama family of models. The company introduced Llama 4 Scout and Llama 4 Maverick in April 2025, touting them as “the first open-weight natively multimodal models with unprecedented context length support.” These models represent a significant advancement in Meta’s AI strategy, featuring a mixture-of-experts (MoE) architecture.

Strategic Pivots Amid Performance Concerns

The timing of these departures is particularly troubling as Meta grapples with delays in rolling out its flagship AI model, Llama 4 Behemoth. According to reporting from The Wall Street Journal highlighted by TechCrunch, the company has postponed the launch due to concerns that the model underperforms on key benchmarks. This setback comes despite Meta’s earlier claims that Behemoth would outperform competing models from OpenAI, Anthropic, and Google on specialized STEM benchmarks.

Meta’s AI ambitions extend beyond mere technical achievements. The company projected that its generative AI products would generate between $2 billion and $3 billion in revenue in 2025, with long-term forecasts ranging from $460 billion to $1.4 trillion by 2035. These lofty financial goals now face additional hurdles as the brain trust behind Llama continues to disperse.

Competitive Pressures in the Open Model Space

In an attempt to maintain momentum, Meta recently launched “Llama for Startups,” a program designed to incentivize startups to adopt its AI models. As reported by TechCrunch, the initiative offers eligible U.S.-based companies “direct support” from Meta’s Llama team and potential funding of up to $6,000 per month for six months to offset development costs.

This push comes as Meta faces increasing competition in the open model space from rivals including DeepSeek, Google, and Alibaba’s Qwen. Despite claiming over a billion downloads for its Llama models, the company’s strategy to establish a dominant model ecosystem appears increasingly vulnerable.

Technological Advancements Amid Organizational Challenges

Meta’s technical roadmap for Llama remains ambitious despite these setbacks. At LlamaCon 2025, the company unveiled plans for Llama 4 that emphasized faster performance, multilingual capabilities across 200 languages, and vastly expanded context windows. As TechNewsWorld reports, these enhancements could theoretically “democratize access to AI on a truly global scale.”

The Llama 4 models already released—Scout and Maverick—feature 17 billion active parameters with varying numbers of experts (16 and 128, respectively). Meta designed Scout to fit on a single H100 GPU with Int4 quantization, while Maverick requires a complete H100 host, demonstrating the company’s focus on both performance and accessibility.

Despite these technological advances, Meta’s AI division has weathered additional controversies in recent months. In April, the company defended itself against allegations of cheating on the LM Arena benchmark by using an optimized version of its Llama 4 Maverick model for testing while releasing a different version to the public.

As the AI landscape continues to evolve rapidly, Meta’s ability to stem the tide of departing talent and deliver on its ambitious roadmap will likely determine whether its open-source AI strategy can successfully challenge the closed ecosystems of competitors like OpenAI and Anthropic.