A Landmark Moment in Open-Source AI
On April 5, 2025, Meta released Llama 4, marking what would become recognized as one of the most significant developments in open-source artificial intelligence. The release represented a dramatic architectural departure from previous Llama models, introducing Mixture of Experts (MoE) architecture to Meta’s flagship open-source family for the first time.
The Technical Leap Forward
Meta unveiled three models under the Llama 4 banner: Scout, Maverick, and Behemoth (with Behemoth still in training at the time of announcement). According to TechCrunch’s coverage, both Scout and Maverick utilized 17 billion active parameters but differed dramatically in their expert configurations and capabilities.
Scout, the smaller configuration, deployed 16 experts with a total parameter count of 109 billion and supported a remarkable 10 million token context window. Maverick scaled up significantly with 128 experts, reaching 400 billion total parameters while maintaining a 1 million token context window. These specifications represented a fundamental shift from the dense transformer architecture that characterized Llama 3 and earlier models.
The training data scale was equally impressive. Meta trained Llama 4 on over 30 trillion tokens—double the amount used for Llama 3, according to the company’s official announcement. This massive training corpus enabled the models to achieve native multimodal capabilities, accepting both text and image inputs while generating text outputs. The models also supported 12 languages, significantly expanding Llama’s multilingual reach.
Why Mixture of Experts Mattered
The adoption of MoE architecture represented a strategic alignment with industry trends. By April 2025, several leading AI labs had demonstrated the efficiency advantages of MoE systems, which activate only a subset of parameters for each inference request. This approach promised both computational efficiency and the ability to scale total parameters without proportionally increasing inference costs.
For the open-source community, Meta’s embrace of MoE was particularly significant. It democratized access to an architecture that had previously been limited to proprietary models, potentially enabling researchers and developers with modest computational resources to work with models boasting hundreds of billions of parameters.
The Licensing Controversy
While Meta positioned Llama 4 as “open source,” the release sparked immediate debate about the terminology. The models were released with “open weights” under a license that imposed a 700 million monthly active user threshold, beyond which organizations required a commercial license from Meta. According to Meta AI’s official documentation, this represented a continuation of the licensing approach from Llama 3.
More controversially, Meta explicitly prohibited users in the European Union from accessing or using Llama 4, citing the region’s AI and privacy regulations. This geographic restriction marked a notable departure from truly open-source principles and generated significant discussion within the AI community during the week following the release.
Mark Zuckerberg’s Open Source Vision
Meta CEO Mark Zuckerberg emphasized the company’s commitment to open-source AI in statements accompanying the release. His positioning of Llama 4 reflected Meta’s broader strategy of building an ecosystem around its AI models, contrasting with the closed approaches of competitors like OpenAI and Anthropic.
The timing of the release, coming roughly one year after Llama 3, demonstrated Meta’s sustained investment in maintaining competitive open-source alternatives to proprietary models.
Industry Context and Competitive Landscape
In early April 2025, the AI landscape was characterized by intense competition between open and closed model approaches. Meta’s release came at a moment when questions about the long-term viability of open-source AI were prominent in industry discussions. The company’s ability to deliver state-of-the-art capabilities while maintaining its open weights approach was seen as a validation of the open-source strategy.
The multimodal capabilities positioned Llama 4 to compete with models like GPT-4V and Claude 3, which had established multimodal processing as a baseline expectation for frontier AI systems.
Immediate Reception
During the week following the April 5 announcement, coverage from major technology publications highlighted both the technical achievements and the licensing complexities. TechCrunch’s reporting noted the scale of the models and the significance of the MoE architecture, while also documenting the EU restriction controversy.
The release generated substantial interest among AI developers and researchers, though comprehensive independent benchmarks were still emerging during the initial coverage period through April 12, 2025.
Historical Significance
Llama 4’s release represented a clear inflection point in open-source AI development. By bringing MoE architecture, massive context windows, and native multimodality to the open weights community, Meta significantly raised the bar for what open-source models could achieve. Whether the licensing terms truly constituted “open source” remained a subject of debate, but the technical capabilities were undeniable.
The event would be remembered as a moment when the gap between proprietary and open-source AI capabilities narrowed considerably, reshaping expectations for what would be possible without exclusive access to closed systems.