Meta has released a new collection of AI models<\/a>, Llama 4, in its Llama family \u2014 on a Saturday, no less. <\/p>\n There are four new models in total: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. All were trained on \u201clarge amounts of unlabeled text, image, and video data\u201d to give them \u201cbroad visual understanding,\u201d Meta says. <\/p>\n The success of open models from Chinese AI lab DeepSeek<\/a>, which perform on par or better than Meta\u2019s previous flagship Llama models, reportedly kicked Llama development into overdrive. Meta is said to have scrambled war rooms to decipher how DeepSeek lowered the cost of running and deploying models like R1<\/a> and V3<\/a>.<\/p>\n Scout and Maverick are openly available on Llama.com<\/a> and from Meta\u2019s partners, including the AI dev platform Hugging Face, while Behemoth is still in training. Meta says that Meta AI, its AI-powered assistant across apps including WhatsApp, Messenger, and Instagram, has been updated to use Llama 4 in 40 countries. Multimodal features are limited to the U.S. in English for now.<\/p>\n Some developers may take issue with the Llama 4 license. <\/p>\n Users and companies \u201cdomiciled\u201d or with a \u201cprincipal place of business\u201d in the EU are prohibited from using or distributing the models<\/a>, likely the result of governance requirements imposed by the region\u2019s AI and data privacy laws. (In the past, Meta has decried these laws<\/a> as overly burdensome.) In addition, as with previous Llama releases, companies with more than 700 million monthly active users must request a special license from Meta, which Meta can grant or deny at its sole discretion.<\/p>\n \u201cThese Llama 4 models mark the beginning of a new era for the Llama ecosystem,\u201d Meta wrote in a blog post<\/a>. \u201cThis is just the beginning for the Llama 4 collection.\u201d <\/p>\n Meta says that Llama 4 is its first cohort of models to use a mixture of experts (MoE) architecture, which is more computationally efficient for training and answering queries. MoE architectures basically break down data processing tasks into subtasks and then delegate them to smaller, specialized \u201cexpert\u201d models.\u00a0<\/p>\n Maverick, for example, has 400 billion total parameters, but only 17 billion active<\/em> parameters across 128 \u201cexperts.\u201d (Parameters roughly correspond to a model\u2019s problem-solving skills.) Scout has 17 billion active parameters, 16 experts, and 109 billion total parameters.<\/p>\n According to Meta\u2019s internal testing, Maverick, which the company says is best for \u201cgeneral assistant and chat\u201d use cases like creative writing, exceeds models such as OpenAI\u2019s GPT-4o<\/a> and Google\u2019s Gemini 2.0<\/a> on certain coding, reasoning, multilingual, long-context, and image benchmarks. However, Maverick doesn\u2019t quite measure up to more capable recent models like Google\u2019s Gemini 2.5 Pro<\/a>, Anthropic\u2019s Claude 3.7 Sonnet<\/a>, and OpenAI\u2019s GPT-4.5<\/a>.<\/p>\n Scout\u2019s strengths lie in tasks like document summarization and reasoning over large codebases. Uniquely, it has a very large context window: 10 million tokens. (\u201cTokens\u201d represent bits of raw text \u2014 e.g. the word \u201cfantastic\u201d split into \u201cfan,\u201d \u201ctas\u201d and \u201ctic.\u201d) In plain English, Scout can take in images and up to millions of words, allowing it to process and work with extremely lengthy documents. <\/p>\n Scout can run on a single Nvidia H100 GPU, while Maverick requires an Nvidia H100 DGX system or equivalent, according to Meta\u2019s calculations.<\/p>\n Meta\u2019s unreleased Behemoth will need even beefier hardware. According to the company, Behemoth has 288 billion active parameters, 16 experts, and nearly two trillion total parameters. Meta\u2019s internal benchmarking has Behemoth outperforming GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro (but not 2.5 Pro) on several evaluations measuring STEM skills like math problem solving. <\/p>\n Of note, none of the Llama 4 models is a proper \u201creasoning\u201d model along the lines of OpenAI\u2019s o1<\/a> and o3-mini<\/a>. Reasoning models fact-check their answers and generally respond to questions more reliably, but as a consequence take longer than traditional, \u201cnon-reasoning\u201d models to deliver answers.<\/p>\n Interestingly, Meta says that it tuned all of its Llama 4 models to refuse to answer \u201ccontentious\u201d questions less often. According to the company, Llama 4 responds to \u201cdebated\u201d political and social topics that the previous crop of Llama models wouldn\u2019t. In addition, the company says, Llama 4 is \u201cdramatically more balanced\u201d with which prompts it flat-out won\u2019t entertain.<\/p>\n \u201c[Y]ou can count on [Lllama 4] to provide helpful, factual responses without judgment,\u201d a Meta spokesperson told TechCrunch. \u201c[W]e\u2019re continuing to make Llama more responsive so that it answers more questions, can respond to a variety of different viewpoints [\u2026] and doesn\u2019t favor some views over others.\u201d<\/p>\n Those tweaks come as some White House allies accuse AI chatbots of being too politically \u201cwoke.\u201d <\/p>\n Many of President Donald Trump\u2019s close confidants, including billionaire Elon Musk and crypto and AI \u201cczar\u201d David Sacks, have alleged that popular AI chatbots\u00a0censor conservative views<\/a>. Sacks has historically\u00a0singled out<\/a>\u00a0OpenAI\u2019s ChatGPT as \u201cprogrammed to be woke\u201d and untruthful about political subject matter.<\/p>\n In actuality, bias in AI is an intractable technical problem. Musk\u2019s own AI company, xAI, has\u00a0struggled<\/a>\u00a0to create a chatbot that doesn\u2019t endorse some political views over others.<\/p>\n That hasn\u2019t stopped companies including OpenAI from adjusting<\/a> their AI models to answer more questions than they would have previously, in particular questions relating to controversial subjects. <\/p>\n<\/div>\n![\"Meta](\"https:\/\/techcrunch.com\/wp-content\/uploads\/2025\/04\/488655517_650996354186993_1043942188415715102_n.png?w=680\")