Meta has officially rolled out the Llama 4 family—four new AI models—marking a significant step forward in its open-source model efforts. Interestingly, the launch happened on a Saturday.
The new lineup includes Llama 4 Scout, Llama 4 Maverick, and the in-development Llama 4 Behemoth. Meta says these models were trained on enormous datasets of unlabeled text, images, and videos, aiming to enhance their overall visual and contextual comprehension.
The rise of competitive open-source models from China’s DeepSeek lab, which matched or even outperformed Meta’s previous generation, is said to have accelerated Llama development. Meta reportedly convened urgent "war rooms" to analyze how DeepSeek managed to cut costs on deploying its R1 and V3 models.
Scout and Maverick are now available via Llama.com and through platforms like Hugging Face. Behemoth, still undergoing training, is yet to be released. Meta has already integrated Llama 4 into Meta AI, its assistant tool across apps like WhatsApp, Messenger, and Instagram, though its multimodal capabilities are currently limited to English in the U.S.
Notably, the Llama 4 licensing includes strict restrictions: users or companies based in the EU are barred from using or redistributing the models—likely a nod to Europe’s complex AI and privacy laws, which Meta has publicly criticized in the past. Also, organizations with more than 700 million monthly users need to apply for a special license, which Meta can choose to grant or deny.
“These Llama 4 models represent the beginning of a new era for the Llama ecosystem,” Meta wrote in a blog post, signaling there’s more to come.
Llama 4 marks the company’s first set of models using a mixture of experts (MoE) architecture—a method designed to improve training and inference efficiency. This structure allows tasks to be split and handled by specialized sub-models.
Maverick, for instance, includes 400 billion total parameters but activates only 17 billion per task across 128 experts. Scout has 17 billion active parameters, 16 experts, and 109 billion total parameters.
Meta’s internal benchmarks suggest Maverick performs better than previous models like GPT-4o and Gemini 2.0 in areas such as coding, reasoning, and multilingual tasks—but it falls short of newer powerhouses like Gemini 2.5 Pro, Claude 3.7 Sonnet, and OpenAI’s GPT-4.5.
Scout is tailored more for document summarization and reasoning over expansive codebases. It also boasts a unique 10-million-token context window, allowing it to process vast amounts of text or multimodal input. Scout can run on a single Nvidia H100 GPU, whereas Maverick requires more robust infrastructure.
Behemoth, the most powerful of the group, is still in training and will demand even greater computing resources. With 288 billion active parameters and nearly two trillion in total, Meta claims Behemoth outperforms GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in math and other STEM-heavy tasks—though Gemini 2.5 Pro remains the frontrunner.
Interestingly, none of the Llama 4 models qualify as full-fledged “reasoning” systems like OpenAI’s o1 or o3-mini, which are designed to fact-check and ensure consistent, accurate outputs—albeit at a slower pace.
Meta also tuned Llama 4 to be more open to controversial or politically sensitive queries. The models now respond more frequently to such topics, whereas earlier Llama versions would often decline to answer. Meta claims these updates make the models more balanced, offering a wider range of responses regardless of political slant.
A company spokesperson told TechCrunch that users “can count on [Llama 4] to provide helpful, factual responses without judgment,” while also handling diverse perspectives more equitably.
These changes arrive amid heightened scrutiny from political circles. Allies of Donald Trump—including Elon Musk and tech investor David Sacks—have accused major AI platforms of leaning left and filtering conservative viewpoints. Sacks has singled out OpenAI’s ChatGPT as “woke” and biased.
But achieving true neutrality in AI is a technically thorny issue. Even Musk’s xAI has struggled to build a chatbot free from perceived political leanings.
Regardless, the push across companies like OpenAI and Meta is clear: make models smarter, more responsive, and more willing to engage with the complex, messy questions users throw their way.
