LONDON: Meta Platforms has released the latest iteration of its large language model (LLM) series, launching Llama 4 Scout and Llama 4 Maverick, which it describes as its “most advanced models yet” and “the best in their class for multimodality.”
In a statement issued over the weekend, the tech giant said both models are part of its broader strategy to build open-source artificial intelligence tools.
The company also announced it is previewing Llama 4 Behemoth, which it called “one of the smartest LLMs in the world and our most powerful yet to serve as a teacher for our new models.”
The Llama 4 collection powers Meta AI, the company’s virtual assistant now embedded in Meta’s suite of platforms including WhatsApp, Messenger, Instagram, and the web.
Both Scout and Maverick models are available for download via Meta’s own repositories and open-source AI platform Hugging Face.
Meta emphasised that the new models are “multimodal,” capable of interpreting and generating across various data types — including text, images, video, and audio.
According to the company, Llama 4 Scout features a 10-million-token context window — a measure of the model’s working memory — and can outperform Google’s Gemma 3 and Gemini 2.0 Flash-Lite, as well as open-source rival Mistral 3.1, “across a broad range of widely reported benchmarks,” while still being compact enough to run on a single Nvidia H100 GPU.
The larger Llama 4 Maverick model is designed to compete with OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash.
Meta claims Maverick delivers comparable results to DeepSeek-V3 in coding and reasoning tasks using “less than half the active parameters.”
The forthcoming Llama 4 Behemoth is reportedly under continued training. Meta stated that Behemoth will contain 2 trillion parameters in total, with 288 billion active at any given time.
The model, which employs a “mixture of experts” (MoE) architecture to optimise resource use, is said to outperform GPT-4.5 and Claude Sonnet 3.7 “on several STEM benchmarks.”
Meta CEO Mark Zuckerberg has touted Behemoth as “the highest performing base model in the world.”
The company plans to share further technical details and product announcements during its upcoming developer event, LlamaCon, scheduled for 29 April.
However, Llama 4’s development did not come without challenges.
According to a report by The Information, Meta delayed the release due to initial performance issues in reasoning and maths, and concerns that its models were less capable than OpenAI’s in generating humanlike voice conversations.
Despite being marketed as open source, Llama 4 has faced criticism over licensing restrictions. The Llama 4 licence prohibits commercial use by companies with more than 700 million monthly active users without prior permission from Meta.
In a 2023 statement, the Open Source Initiative noted that such restrictions “take it out of the category of ‘Open Source.’”
The launch of Llama 4 comes amid increasing investment in AI infrastructure by major tech companies.
Meta has reportedly earmarked up to $65 billion for AI-related expansion this year, as investors increasingly demand visible returns on their AI bets in the wake of OpenAI’s breakthrough with ChatGPT.
