Meta unveiled a new era of its most advanced and latest LLM multimodal AI, the Llama 4, which marks a significant leap in Meta’s artificial intelligence capabilities in the global AI race. During the Launch on April 5, Meta introduced the first models in the Llama 4 line designed to excel in performance, efficiency, accessibility, and the ability to give the users a more personalized AI experience.
Previous models of Llama were basically LLMs (large language models) trained to perform tasks based on natural language processing, like text generation, translation, summarization, question answering, and text completion. The new multimodal model is an extension of an LLM that is capable of processing and understanding information from multiple modalities like images, audio, video, sensor data, physiological signals, or structured data.
The launch introduces three new models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. Llama 4 Scout has 17 billion active parameters with 16 experts (sub-networks), which helps increase the model’s capacity and efficiency, making it outperform other multimodal models with similar parameters and more powerful than previous Llama models.
The Llama 4 Scout has a single NVIDIA H100 GPU (graphics processing unit) and offers a 10 million token context window, which allows the model to understand and generate coherent responses based on much longer and complex inputs. Features like this enable the model to deliver better results than other multimodal AIs like Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1.
Llama 4 Maverick has 17 billion active parameters and 128 experts, which allows for a larger model capacity without an increase in computational cost. This gives the model an edge over GPT-4o and Gemini 2.0 Flash across a broad range of evaluations and sets it on the same capacity as DeepSeek v3 on reasoning and coding, achieving such feats at less than half the active parameters.
The third model is a “teacher” model, designed to train other new models in the Llama 4 line and the most powerful Llama AI model. Llama 4 Behemoth has 288 billion active parameters with 16 experts, making it part of the top percentile of the world’s smartest LLMs.
Behemoth outperformed Claude Sonnet 3.7, GPT-4.5 and Gemini 2.0 pro in STEM (science, technology, engineering, and mathematics)-focused benchmarks, like the MATH-500 test for mathematical skills and GPQA Diamond.
Behemoth as it is still in training and has not been released by Meta. Maverick and Scout, however, are currently available and can be downloaded from llama.com and Hugging Face. The Llama 4 models are also accessible via WhatsApp Messenger, Instagram Direct, and Meta’s website.
Currently, the AI market is full of intense competition between nations and tech giants, all in a bid to achieve AI supremacy. China and the U.S. remain the forerunners as they have the most significant investments and developments in the industry.
The launch of Llama 4 by Meta is believed by experts to give the U.S. a stronger edge over its counterpart, especially as the model’s license allows it to be readily accessible to developers and businesses.
Meta has achieved a significant milestone as it unveils Llama 4 by pushing the boundaries of the known variables of AI performance and multimodality. Through democratizing access to the models, Meta is positioning itself in the position of becoming a key player in shaping the future of artificial intelligence.
