Meta has unveiled its latest suite of AI models, Llama 4, which now powers the Meta AI assistant across its platforms, including WhatsApp, Messenger, and Instagram. The new release features two models available for download: Llama 4 Scout, which is compact enough to operate on a single Nvidia H100 GPU, and Llama 4 Maverick, designed to be comparable to GPT-4o and Gemini 2.0 Flash. Additionally, Meta CEO Mark Zuckerberg announced that the Llama 4 Behemoth model is still under training and is touted as the “highest performing base model in the world.”
Meta claims that Llama 4 Scout boasts a context window of 10 million tokens, outperforming Google’s Gemini 3 and Gemini 2.0 Flash-Lite models, as well as the open-source Mistral 3.1, across various benchmark tests while still maintaining its compact design. The company presents similar performance comparisons for the larger Maverick model against OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash, asserting that its efficiency in coding and reasoning tasks is on par with DeepSeek-V3 while utilizing fewer active parameters.
The upcoming Llama 4 Behemoth is expected to contain a staggering 288 billion active parameters, totaling 2 trillion parameters overall. Although not yet released, Meta anticipates that Behemoth will outperform competitors like GPT-4.5 and Claude Sonnet 3.7 on various STEM-related benchmarks.
Meta has adopted a “mixture of experts” (MoE) architecture for the Llama 4 series, allowing the models to utilize only the necessary segments for specific tasks, thus conserving computational resources. Further details regarding Meta’s AI models and future product development are expected to be revealed at the upcoming LlamaCon conference on April 29th.
Consistent with previous offerings, Meta describes the Llama 4 collection as “open-source.” However, the licensing terms have faced criticism. Specifically, the Llama 4 license mandates that commercial entities with over 700 million monthly active users obtain permission from Meta prior to utilizing its models. This stipulation has led the Open Source Initiative to assert that the licensing does not meet open-source criteria.
