Meta Introduces Llama 4: Next-Gen Open Source AI Models

04/05/2025 Artificial Intelligence

Meta has just launched its latest AI models, the Llama 4 family, marking a significant step in the open-source AI landscape. This new collection includes Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. These models were trained on a massive dataset of unlabeled text, image, and video data to achieve **broad visual understanding**.

The development of Llama 4 was reportedly accelerated due to the success of open models from the Chinese AI lab DeepSeek, which outperformed Meta's previous Llama models. Meta responded by intensifying efforts to understand how DeepSeek managed to reduce the costs associated with running and deploying models like R1 and V3.

Llama 4 Scout and Maverick are readily available on Llama.com and through Meta's partners like Hugging Face. However, Behemoth is still undergoing training. Meta AI, the AI-powered assistant integrated into apps like WhatsApp, Messenger, and Instagram, has already been updated with Llama 4 in 40 countries, although multimodal features are currently limited to the U.S. in English.

Licensing Restrictions

Some developers might find the Llama 4 license restrictive. Entities based in the EU are prohibited from using or distributing the models, likely due to the region's stringent AI and data privacy laws. Additionally, similar to previous Llama releases, companies with over 700 million monthly active users must request a special license from Meta, which Meta can approve or deny at its discretion.

Meta emphasizes that the Llama 4 models represent the beginning of a new era for the Llama ecosystem. These models are the first to utilize a mixture of experts (MoE) architecture, enhancing computational efficiency for training and query responses. The MoE architecture divides data processing into subtasks, assigning them to smaller, specialized “expert” models.

For instance, Maverick boasts 400 billion total parameters, but only 17 billion active parameters across 128 experts. Scout has 17 billion active parameters, 16 experts, and 109 billion total parameters.

Performance and Capabilities

Meta's internal testing suggests that Maverick, designed for general assistant and chat use cases, outperforms models like OpenAI's GPT-4o and Google's Gemini 2.0 in coding, reasoning, multilingual tasks, long-context understanding, and image benchmarks. However, it falls short of models like Google's Gemini 2.5 Pro, Anthropic's Claude 3.7 Sonnet, and OpenAI's GPT-4.5.

Scout excels in document summarization and reasoning over large codebases. Its standout feature is its massive context window of 10 million tokens, enabling it to process and work with extremely lengthy documents. Scout can operate on a single Nvidia H100 GPU, while Maverick requires an Nvidia H100 DGX system or equivalent.

Behemoth, Meta's unreleased model, demands even more powerful hardware, featuring 288 billion active parameters, 16 experts, and nearly two trillion total parameters. Meta's benchmarks show that Behemoth outperforms GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in STEM skills evaluations, such as math problem-solving.

It's worth noting that none of the Llama 4 models are true “reasoning” models like OpenAI’s o1 and o3-mini. Reasoning models prioritize fact-checking, leading to more reliable responses but longer processing times.

Addressing Bias and Contentious Questions

Meta has also tuned Llama 4 to be more willing to answer “contentious” questions. The models now respond to debated political and social topics that previous Llama models avoided. Meta claims that Llama 4 is more balanced in handling prompts that it previously wouldn’t entertain.

According to a Meta spokesperson, Llama 4 provides helpful, factual responses without judgment and is more responsive to diverse viewpoints. These adjustments come amid accusations from White House allies that AI chatbots are politically biased.

Bias in AI remains a complex technical challenge, with companies like OpenAI and xAI constantly working to create models that don't disproportionately favor certain political views.