AIFEATUREDLatestTOP STORIESWeb Development

Meta Unveils Llama 4 Scout and Maverick


Meta Platforms has introduced two new iterations of its large language model (LLM), Llama 4 Scout and Llama 4 Maverick, released on April 5, 2025. These models represent a significant development in multimodal AI systems, capable of processing text and images. Available as open-weight models under Meta’s own licence (not an open-source licence in the strict sense), they are positioned as Meta’s most advanced offerings to date, particularly for multimodal applications. Let’s examine the specifications, implications, and accessibility of these models…

Update (10 November 2026): The largest Llama 4 model mentioned below, Behemoth, was not released alongside Scout and Maverick; in May 2025 the Wall Street Journal reported that Meta was delaying its rollout because engineers were struggling to improve it significantly.

The Llama 4 models differ from predecessors by integrating multimodal AI capabilities, moving beyond the text-only focus of models like OpenAI’s GPT-4. According to Meta’s official announcement, Llama 4 Scout and Maverick are natively multimodal, built to understand text and images together for tasks such as image understanding and visual reasoning. Meta also said that Meta AI, built with Llama 4, is available in WhatsApp, Messenger and Instagram Direct.

Technically, Llama 4 Scout features 17 billion active parameters (16 experts) and a 10-million-token context window, optimised for efficiency on a single Nvidia H100 GPU. It suits developers managing large datasets or complex code. Conversely, Llama 4 Maverick operates with 400 billion total parameters across 128 specialised components, excelling in conversational and creative functions. Meta’s blog post says Maverick exceeds comparable models such as GPT-4o and Google’s Gemini 2.0 on coding, reasoning, multilingual, long-context and image benchmarks, at a lower cost than Llama 3.3 70B; these are Meta’s own benchmark claims. The models were distilled from Llama 4 Behemoth, a much larger model still in training at the time of release.

A key aspect of this release is its open-weight availability. The models can be downloaded from llama.com and Hugging Face. This aligns with Meta’s strategy to promote innovation, allowing developers to customise them for applications like chatbots or analysis of Zoom recordings. The Meta AI platform was updated to use Llama 4 in 40 countries at launch, according to TechCrunch, with multimodal features initially limited to English.

This open-weight approach contrasts with proprietary models from OpenAI and Google. Meta’s CEO, Mark Zuckerberg, underscored this direction in a statement on Instagram, reflecting the company’s plan to spend $60 billion to $65 billion on capital expenditure in 2025, much of it on AI infrastructure. The Llama 4 release follows industry trends set by ChatGPT, intensifying competition in AI development.

However, access is not universal. EU developers face restrictions due to AI regulations, and companies with more than 700 million monthly users must request a special licence from Meta. For eligible users, the models are available for immediate download from the official page, supporting projects ranging from smart assistants to custom AI solutions.

Meta planned to discuss these advancements at LlamaCon on April 29, 2025. For developers and businesses leveraging Nvidia hardware, Llama 4 Scout and Maverick provide a robust foundation for exploring multimodal AI. This release marks a notable step in Meta’s AI strategy, with implications for the broader technology landscape.


We’d love your comments on today’s topic!

For more articles like this one, click here.

Thought for the day:

“Success in creating AI would be the biggest event in human history. Unfortunately, it might also be the last, unless we learn how to avoid the risks.”

Stephen Hawking

Leave a Reply

Your email address will not be published. Required fields are marked *