Scout:\u00a017 billion active parameters, 109 billion total parameters, and a context window of 10 million tokens.\u00a0<\/p>\n
Maverick:\u00a017 billion active parameters, 400 billion total parameters, and a context window of 1 million tokens.\u00a0<\/p>\n
Behemoth:\u00a0Not yet released\u00a0but\u00a0will\u00a0have 288 billion active\u00a0parameters\u00a0and 2 trillion total parameters.\u00a0\u00a0<\/p>\n
(In data science, tokens are subdivided bits of raw data, like the syllables \u201cfan,\u201d \u201ctas\u201d and \u201ctic\u201d in the word \u201cfantastic.\u201d)\u00a0\u00a0<\/p>\n
A model\u2019s context, or context window, refers to input data (e.g., text) that the model considers before generating output (e.g.,\u00a0additional\u00a0text). Long context can prevent models from \u201cforgetting\u201d the content of recent docs and data, and from veering off topic and extrapolating wrongly. However, longer context windows can also\u00a0result in the model \u201cforgetting\u201d certain safety guardrails<\/a>\u00a0and being more prone to produce content that is in line with the conversation, which has led\u00a0some users toward\u00a0delusional thinking<\/a>.\u00a0\u00a0<\/p>\n For reference,\u00a0the\u00a010 million context window\u00a0that Llama 4 Scout promises\u00a0roughly equals\u00a0the text of about 80 average novels.\u00a0Llama 4\u00a0Maverick\u2019<\/a>s\u00a01 million context window equals about eight novels.\u00a0\u00a0<\/p>\n Techcrunch event<\/p>\n \n\t\t\t\t\t\t\t\t\tSan Francisco All\u00a0of\u00a0the Llama 4 models\u00a0were trained on \u201clarge amounts of unlabeled text, image, and video data\u201d to give them \u201cbroad visual understanding,\u201d\u00a0as well as on 200 languages,\u00a0according to Meta.\u00a0\u00a0<\/p>\n Llama 4 Scout and Maverick are Meta\u2019s first open-weight natively multimodal models.\u00a0They\u2019re\u00a0built using a \u201cmixture-of-experts\u201d (MoE) architecture, which reduces computational load\u00a0and improves efficiency in training and inference. Scout, for example, has 16 experts, and Maverick has 128 experts.\u00a0\u00a0\u00a0<\/p>\n Llama 4 Behemoth includes 16 experts, and Meta is referring to it as a teacher for the smaller models.\u00a0<\/p>\n Llama 4 builds on the Llama 3 series, which included 3.1 and 3.2 models widely used for instruction-tuned applications and cloud deployment.\u00a0<\/p>\n What can Llama do? <\/p>\n Like other generative AI models, Llama can perform a range of different assistive tasks, like coding and answering basic math questions, as well as summarizing documents in\u00a0at least 12\u00a0languages (Arabic,\u00a0English, German, French,\u00a0Hindi, Indonesian,\u00a0Italian, Portuguese, Hindi,\u00a0Spanish, Tagalog, Thai, and Vietnamese). Most text-based workloads \u2014 think analyzing\u00a0large\u00a0files like PDFs and spreadsheets \u2014 are within its purview, and all Llama 4\u00a0models\u00a0support text, image, and video input.\u00a0<\/p>\n Llama 4 Scout is designed for longer workflows and massive data analysis. Maverick is a generalist model that is better at balancing reasoning power and response speed, and is suitable for coding, chatbots, and technical assistants. And Behemoth is designed for advanced research, model distillation, and STEM tasks.\u00a0\u00a0<\/p>\n Llama models, including Llama 3.1, can be configured to\u00a0leverage\u00a0third-party applications, tools, and APIs to perform tasks. They are trained to use Brave Search for answering questions about recent events;\u00a0the Wolfram Alpha API for math- and science-related queries;\u00a0and a Python interpreter for validating code. However, these tools require\u00a0proper configuration and are not automatically enabled out of the box.\u00a0<\/p>\n Where can I use Llama?<\/p>\n If\u00a0you\u2019re\u00a0looking to simply chat with Llama,\u00a0it\u2019s\u00a0powering the Meta AI chatbot experience<\/a>\u00a0on Facebook Messenger, WhatsApp, Instagram, Oculus,\u00a0and Meta.ai\u00a0in 40 countries. Fine-tuned versions of Llama are used in Meta AI experiences in over 200 countries and territories.\u00a0\u00a0<\/p>\n Llama 4 models Scout and Maverick are available on Llama.com and Meta\u2019s partners, including the AI developer platform Hugging Face. Behemoth is still in training.\u00a0Developers\u00a0building with\u00a0Llama can download,\u00a0use,\u00a0or fine-tune the model across most of the popular cloud platforms.\u00a0Meta claims it has\u00a0more than\u00a025 partners hosting Llama, including Nvidia, Databricks,\u00a0Groq, Dell,\u00a0and Snowflake.\u00a0And while \u201cselling access\u201d to Meta\u2019s openly available models\u00a0isn\u2019t\u00a0Meta\u2019s business model, the company makes some money\u00a0through\u00a0revenue-sharing agreements<\/a>\u00a0with model hosts.\u00a0<\/p>\n Some of these partners have built\u00a0additional\u00a0tools and services on top of Llama, including tools that let the models reference proprietary data and enable them to run at lower latencies.\u00a0<\/p>\n Importantly, the Llama license\u00a0constrains how developers can deploy the model<\/a>: App developers with more than 700 million monthly users must request a special license from Meta that the company will grant on its discretion.\u00a0<\/p>\n In May 2025, Meta launched a\u00a0new program<\/a>\u00a0to incentivize startups to adopt its Llama models. Llama for Startups gives companies support from Meta\u2019s Llama team and access to potential funding.\u00a0\u00a0<\/p>\n Alongside Llama, Meta provides tools intended to make the model \u201csafer\u201d to use:\u00a0\u00a0<\/p>\n Llama Guard, a moderation framework.\u00a0<\/p>\n CyberSecEval, a cybersecurity risk assessment suite.\u00a0<\/p>\n Llama Firewall, a security guardrail designed to enable building secure AI systems.\u00a0<\/p>\n Code Shield, which provides support for inference-time filtering of insecure code produced by LLMs.\u00a0\u00a0<\/p>\n Llama\u00a0Guard tries to detect potentially problematic content either fed into \u2014 or generated \u2014 by a Llama model, including content relating to criminal activity, child exploitation, copyright violations, hate, self-harm and sexual abuse.\u00a0That said,\u00a0it\u2019s\u00a0clearly not a silver bullet since\u00a0Meta\u2019s own previous guidelines<\/a>\u00a0allowed the chatbot to engage in sensual and romantic chats with minors, and some reports show those\u00a0turned\u00a0into\u00a0sexual conversations<\/a>.\u00a0Developers can\u00a0customize<\/a>\u00a0the categories of blocked content and apply the blocks to all the languages Llama supports.\u00a0<\/p>\n Like Llama Guard, Prompt Guard can block text intended for Llama, but only text meant to \u201cattack\u201d the model and get it to behave in undesirable ways. Meta claims that\u00a0Llama\u00a0Guard can defend against explicitly malicious prompts (i.e., jailbreaks that\u00a0attempt\u00a0to get around Llama\u2019s built-in safety filters) in addition to prompts that\u00a0contain\u00a0\u201cinjected inputs<\/a>.\u201d\u00a0The Llama Firewall works to detect and prevent risks like prompt injection, insecure code, and risky tool interactions. And Code Shield helps mitigate insecure code suggestions and offers secure command execution for seven programming languages.\u00a0<\/p>\n As for\u00a0CyberSecEval,\u00a0it\u2019s\u00a0less a tool than a collection of benchmarks to measure model security.\u00a0CyberSecEval\u00a0can assess the risk a Llama model poses (at least according to Meta\u2019s criteria) to app developers and end users in areas like \u201cautomated social engineering\u201d and \u201cscaling offensive cyber operations.\u201d\u00a0<\/p>\n Llama\u2019s limitations<\/p>\n Llama comes with certain risks and limitations, like all generative AI models.\u00a0For example, while its most recent model has multimodal features, those are\u00a0mainly limited\u00a0to\u00a0the English\u00a0language\u00a0for now.\u00a0<\/p>\n Zooming out,\u00a0Meta used a dataset of pirated e-books<\/a>\u00a0and articles to train its Llama models. A federal judge recently sided with Meta in a copyright lawsuit brought against the company by 13 book authors, ruling that\u00a0the use of copyrighted works for training fell under \u201cfair use.\u201d However, if Llama\u00a0regurgitates<\/a>\u00a0a copyrighted snippet\u00a0and someone uses it in a product, they could potentially be infringing on copyright and be liable.\u00a0\u00a0<\/p>\n Meta also\u00a0controversially trains its AI on Instagram and Facebook posts,\u00a0photos\u00a0and captions, and\u00a0makes it difficult for users to opt out<\/a>.\u00a0\u00a0<\/p>\n Programming is another area where\u00a0it\u2019s\u00a0wise to tread lightly when using Llama.\u00a0That\u2019s\u00a0because Llama might \u2014\u00a0perhaps\u00a0more\u00a0so than\u00a0its generative AI counterparts \u2014\u00a0produce buggy or insecure code<\/a>.\u00a0On\u00a0LiveCodeBench<\/a>, a\u00a0benchmark<\/a>\u00a0that tests AI models on competitive coding problems, Meta\u2019s Llama 4 Maverick model achieved a score of 40%.\u00a0That\u2019s\u00a0compared to\u00a085% for OpenAI\u2019s GPT-5 high\u00a0and\u00a083% for\u00a0xAI\u2019s\u00a0Grok 4 Fast.\u00a0<\/p>\n As always,\u00a0it\u2019s\u00a0best to have a human expert review any AI-generated code before incorporating it into a service or software.\u00a0<\/p>\n
\n\t\t\t\t\t\t\t\t\t\t\t\t\t|
\n\t\t\t\t\t\t\t\t\t\t\t\t\tOctober 27-29, 2025\n\t\t\t\t\t\t\t<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2025\/10\/Llama-4-coding-benchmark.png\")
Image Credits:Artificial Analysis<\/p>\n