The peculiarity of LLAMA 4 models lies in the unprecedented magnitude of MOE implementation, with up to 128 experts for Maverick. Meta also applies a technique called “early fusion” which deals with text, images and videos as a unique tokens sequence from the start of the process, allowing a real native multimodal understanding.
On the Benchmark Lmarena, Llama 4 Maverick would reach an Elo score of 1417, positioning itself in the most advanced models on the market. As for the Behemoth model, again in the training phase with its 288 billion active parameters and its 16 experts totaling nearly 2 billions of parameters, Meta says that it already surpasses GPT-4.5, Claude Sonnet 3.7 and Gemini 2.0 Pro on several scientific benchmarks. These new models are already available on the Meta AI platform as well as via WhatsApp, Messenger and Direct Instagram in 40 countries. However, multimodal features are currently limited to English-speaking users in the United States.
