3 min read
Meta has officially launched its latest open-source AI models — Llama 4 Scout and Llama 4 Maverick — under its Llama 4 family, with a fourth model, Behemoth, still in training. The announcement was made by Meta CEO Mark Zuckerberg in an Instagram Reel, where he described the release as “a milestone for Meta AI and for open source.”
View this post on Instagram
These models are now integrated into Meta AI across WhatsApp, Messenger, and Instagram Direct, and are available via Meta.ai in 40 countries, though multimodal features are currently limited to the U.S. in English. Both models are also available for free download from Meta and Hugging Face, though they come with licensing restrictions — including a ban on use by companies based in the EU and organizations with over 700 million monthly users without explicit Meta approval.
Llama 4 Scout
Scout is a smaller yet powerful multimodal model with 17 billion active parameters (out of 109 billion total) distributed across 16 expert networks, making it Meta’s most efficient model to date. It supports a staggering context length of 10 million tokens — allowing it to process extremely long documents, codebases, and images. Scout runs efficiently on a single Nvidia H100 GPU and outperforms models like GPT-4o and DeepSeek V3 in multiple benchmarks, especially in reasoning and summarization tasks.
Llama 4 Maverick
Described by Meta as the “workhorse,” Maverick is built for general assistance and creative writing tasks. It consists of 400 billion parameters and utilizes 128 experts, with only 17 billion active at any time. It has a context window of 1 million tokens and delivers high performance on coding, multilingual, and image-understanding tasks. With an ELO score of 1,417 on the LMArena leaderboard, Maverick ranks just below Gemini 2.5 Pro and surpasses models such as GPT-4.5 and Grok 3 — all while using a fraction of their active parameters.
Llama 4 Behemoth and Beyond
Llama 4 Behemoth, still in training, boasts nearly 2 trillion total parameters, with 288 billion active across 16 experts. Meta claims it surpasses GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro on STEM-related benchmarks. Although Behemoth is not yet available, its introduction is expected to mark a significant leap in large-scale AI capabilities.
All models in the Llama 4 family are based on a Mixture of Expert (MoE) architecture, which breaks down tasks and routes them to smaller expert models for improved performance and efficiency. This design allows Meta’s models to achieve competitive results with fewer active parameters, enabling cost-effective deployment and inference.
Balancing Responsiveness and Controversy
Meta has also fine-tuned the Llama 4 series to reduce the frequency of refusing “contentious” prompts. According to the company, these models provide more balanced and responsive answers on politically and socially debated topics, striving to offer factual responses “without judgment.” Meta’s approach comes in response to ongoing criticisms regarding political bias in AI models.
“These Llama 4 models mark the beginning of a new era for the Llama ecosystem,” Meta emphasized in a blog post, hinting at the continued expansion and evolution of its AI strategy.
