Meta Unveils Llama 4 with Multimodal AI
Meta has unveiled Llama 4, the next generation of its open-source AI model, featuring deep multimodal capabilities. Llama 4 is designed to natively process and generate content across text, images, and video within a single workflow. This positions it as a powerful, open-source backbone for building integrated creative automation tools.
Llama 4 arrives as a "herd" of three distinct models: the powerful but still-in-development 2-trillion parameter "Behemoth," a single-GPU-optimized "Scout," and the cost-effective "Maverick" model. This release follows the trajectory of Llama 3.1, whose 405B parameter flagship model already outperformed competitors like GPT-4 and Claude 3.5 Sonnet in key benchmarks like translation and summarization. The launch has been met with controversy regarding its benchmark performance, with allegations circulating of "contamination"—training the model on the test questions themselves to inflate scores. Meta's VP of Generative AI, Ahmad Al-Dahle, has publicly denied these claims, stating, "that's simply not true and we would never do that." This new iteration also reignites the debate over Meta's "open-source" classification. The Open Source Initiative (OSI) has previously criticized Meta's licensing, arguing the restrictions on commercial use and lack of full transparency mean the models are not truly open-source. This has led to accusations of "open-washing," or using the term for marketing while maintaining corporate control. Meta's strategy aims to commoditize the AI model layer, encouraging widespread adoption by researchers, developers, and startups to build on their technology. By open-sourcing the model weights, Meta shifts the significant computational and infrastructure costs to those who deploy Llama, differentiating its approach from the API-based, closed-garden models of OpenAI and Google. However, the "open" strategy has its limits; Meta did not release Llama 4 in Europe, citing regulatory difficulties. This decision raises further questions about the universal accessibility of the models and the complexities of navigating global AI governance.
