Meta Releases Llama 4: Next-Gen AI Models

Meta Unveils Llama 4: A New Generation of AI Models

Meta has introduced its latest suite of artificial intelligence models, collectively known as Llama 4, as part of the ongoing Llama family. The release occurred on a Saturday, marking a notable timing for this significant update.

Introducing the Llama 4 Models

The new collection comprises three distinct models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. These models were all trained utilizing extensive datasets of unlabeled text, images, and video content. This training process aims to provide them with a comprehensive understanding of visual information, as stated by Meta.

Reportedly, the rapid advancement of open-source models from DeepSeek, a Chinese AI laboratory, which demonstrated performance comparable to or exceeding Meta’s prior Llama models, spurred accelerated development within the Llama project. Meta is said to have established dedicated teams to analyze how DeepSeek achieved lower costs for running and deploying models such as R1 and V3.

Availability and Integration

Both Scout and Maverick are now publicly accessible through Llama.com and via Meta’s collaborative partners, including the AI development platform Hugging Face. Behemoth, however, remains in the training phase.

Meta has also announced an update to Meta AI, its AI-powered assistant integrated across various applications like WhatsApp, Messenger, and Instagram, incorporating Llama 4 technology. Currently, multimodal capabilities are limited to users in the United States utilizing the English language.

Licensing Restrictions

The licensing terms for Llama 4 may present challenges for some developers.

Entities “domiciled” or maintaining a “principal place of business” within the European Union are restricted from utilizing or distributing these models. This limitation likely stems from the governance requirements imposed by the region’s AI and data privacy regulations. Meta has previously expressed concerns regarding the perceived burdens of these laws.

Furthermore, consistent with previous Llama releases, organizations exceeding 700 million monthly active users are required to obtain a specific license from Meta, which is subject to Meta’s discretionary approval or denial.

Technical Specifications and Architecture

“These Llama 4 models represent a pivotal moment in the evolution of the Llama ecosystem,” Meta declared in a blog post. “This release is merely the initial step for the Llama 4 collection.”

Meta highlights that Llama 4 represents its inaugural group of models employing a mixture of experts (MoE) architecture. This architecture enhances computational efficiency during both training and query processing. Essentially, MoE architectures divide complex data processing into smaller tasks, delegating them to specialized “expert” models.

Model Parameters and Performance

For instance, Maverick boasts 400 billion total parameters, yet only 17 billion active parameters are utilized across 128 “experts.” (Parameters are generally indicative of a model’s problem-solving capacity.) Scout features 17 billion active parameters, 16 experts, and a total of 109 billion parameters.

Based on Meta’s internal evaluations, Maverick, designed for “general assistant and chat” applications like creative writing, surpasses models such as OpenAI’s GPT-4o and Google’s Gemini 2.0 on specific coding, reasoning, multilingual, long-context, and image benchmarks. However, Maverick does not yet achieve the performance levels of more advanced recent models like Google’s Gemini 2.5 Pro, Anthropic’s Claude 3.7 Sonnet, and OpenAI’s GPT-4.5.

Scout excels in tasks such as document summarization and reasoning over extensive codebases. Notably, it possesses a remarkably large context window of 10 million tokens. (“Tokens” are fundamental units of text – for example, the word “fantastic” can be divided into “fan,” “tas,” and “tic.”) In simpler terms, Scout can process images and up to millions of words, enabling it to analyze and work with exceptionally long documents.

Meta estimates that Scout can operate on a single Nvidia H100 GPU, while Maverick necessitates an Nvidia H100 DGX system or equivalent hardware.

Behemoth: The Most Powerful Model

Meta’s forthcoming Behemoth model will demand even more robust hardware. The company specifies that Behemoth has 288 billion active parameters, 16 experts, and nearly 2 trillion total parameters. Internal benchmarking by Meta indicates that Behemoth outperforms GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro (though not 2.5 Pro) on several assessments measuring STEM skills, including mathematical problem-solving.

It is important to note that none of the Llama 4 models are designed as dedicated “reasoning” models comparable to OpenAI’s o1 and o3-mini. Reasoning models prioritize fact-checking and generally provide more reliable responses, albeit at the cost of increased processing time.

Addressing Bias and Contentious Topics

Interestingly, Meta states that it has refined all of its Llama 4 models to reduce their tendency to decline answering “contentious” questions. According to the company, Llama 4 is more responsive to “debated” political and social topics that previous Llama models would have avoided. Additionally, Meta asserts that Llama 4 exhibits “dramatically more balanced” criteria for refusing to address certain prompts.

“[Y]ou can rely on [Llama 4] to deliver helpful, factual responses without bias,” a Meta spokesperson conveyed to TechCrunch. “[W]e’re continually enhancing Llama’s responsiveness so that it answers a wider range of questions, can accommodate diverse viewpoints… and doesn’t favor specific perspectives.”

These adjustments coincide with accusations from some White House allies that AI chatbots exhibit excessive political “bias.”

Numerous close associates of former President Donald Trump, including billionaire Elon Musk and crypto and AI “czar” David Sacks, have alleged that popular AI chatbots censor conservative viewpoints. Sacks has consistently criticized OpenAI’s ChatGPT as “programmed to be woke” and inaccurate regarding political subjects.

In reality, bias in AI represents a complex and persistent technical challenge. Musk’s own AI company, xAI, has encountered difficulties in creating a chatbot that avoids endorsing certain political perspectives.

Despite this, companies, including OpenAI, are adjusting their AI models to address a broader range of questions, particularly those pertaining to controversial subjects.