Meta Unveils Llama 4: A Groundbreaking Advance in AI Model Collection

Meta* has recently launched a new series of artificial intelligence models, Llama 4, from its Llama family, with the announcement made on Saturday. This collection includes four new models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. According to Meta*, these models were trained on extensive datasets consisting of unlabelled text, images, and videos to develop a «comprehensive visual understanding.»

The push to expedite the development of Llama comes in response to the success of open models from the Chinese AI laboratory DeepSeek, which reportedly perform on par with or exceed prior flagship Llama models from Meta. It’s said that Meta assembled dedicated teams to analyze how DeepSeek has lowered the costs associated with deploying models like R1 and V3.

Scout and Maverick are now publicly available on Llama.com and through Meta partners, such as the Hugging Face AI development platform, while Behemoth is still undergoing training. According to Meta, its AI assistant, which operates in apps like WhatsApp, Messenger, and Instagram, has been updated to incorporate Llama 4 in 40 countries. The multimodal features are currently limited to English language users in the U.S.

Some developers may face challenges related to the licensing of Llama 4. Users and organizations «residing» or with a «primary place of business» in the EU are prohibited from utilizing or distributing the models, likely due to regulatory requirements stemming from regional AI and data privacy laws. In the past, Meta has criticized such legislation as overly burdensome. Moreover, companies with over 700 million active users per month have to seek a special license from Meta, which can be granted or denied at their discretion.

Meta noted in a blog post, «These Llama 4 models mark the dawn of a new era for the Llama ecosystem.» They further emphasize that this is merely the beginning of the Llama 4 collection.

According to Meta, Llama 4 represents the first group of models employing a mixture of experts (MoE) architecture, which enhances computational efficiency in training and responding to queries. MoE architectures primarily break data processing tasks into sub-tasks and assign them to smaller specialized «expert» models.

For instance, Maverick has 400 billion parameters but only 17 billion active parameters spread across 128 «experts.» Scout features 17 billion active parameters, 16 experts, and a total of 109 billion parameters.

Internal testing by Meta suggests that Maverick, positioned as «an ideal assistant for general purposes and chatting,» excels in creative writing tasks, outperforming models like OpenAI’s GPT-4o and Google’s Gemini 2.0 in several coding, reasoning, multilingual, long-context, and image-related metrics. However, Maverick falls short compared to newer, more efficient models like Google’s Gemini 2.5 Pro, Anthropic’s Claude 3.7 Sonnet, and OpenAI’s GPT-4.5.

Scout shines in tasks such as document summarization and reasoning over extensive codebases. Notably, it boasts an exceptionally large context window of 10 million tokens, allowing it to process and work with incredibly lengthy documents.

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

The unreleased Behemoth model will require even more robust hardware. With 288 billion active parameters, 16 experts, and nearly two trillion parameters overall, internal benchmarking indicates that Behemoth surpasses GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro (but not 2.5 Pro) in several evaluations of STEM capabilities.

However, it’s important to note that none of the Llama 4 models qualify as full-fledged «reasoning» models comparable to OpenAI’s o1 and o3-mini. Reasoning models verify facts in their responses and typically answer questions more reliably, albeit at a slower pace compared to traditional «non-reasoning» models.

Interestingly, Meta claims to have configured all its Llama 4 models to decrease instances of refusal to engage with «controversial» topics. The company states that Llama 4 provides responses to «discussed» political and social issues that previous Llama models would not address. Additionally, it asserts that Llama 4 is «significantly more balanced» and will not entertain prompts in a frivolous manner.

«You can expect Llama 4 to provide useful, factual answers without bias,» a Meta representative told TechCrunch. «We continue to enhance Llama’s responsiveness to address a wider array of inquiries, engage with diverse viewpoints, and avoid favoritism toward any single perspective.»

These adjustments come amid accusations from some White House allies claiming that AI chatbots are overly «woke» politically.

Many close associates of former President Donald Trump, including billionaire Elon Musk and «crypto king» and AI expert David Sacks, have argued that popular AI chatbots are «censoring conservative viewpoints.» Sacks has historically pointed out that OpenAI’s ChatGPT is «programmed for wokeness» and deceptive regarding political matters.

Bias in AI is, in fact, a complex challenge. Musk’s AI venture, xAI, has been striving to create a chatbot that does not favor one political viewpoint over another. Nevertheless, firms like OpenAI have started to adjust their AI models to handle a broader range of inquiries, particularly those concerning contentious issues.

(Note: Meta* and its products, including Facebook and Instagram, are prohibited in the Russian Federation)

[Source](https://the-decoder.com/google-adds-web-search-capabilities-to-notebooklm/)