Like each Large Tech firm today, Meta has its personal flagship generative AI mannequin, referred to as Llama. Llama is considerably distinctive amongst main fashions in that it’s “open,” which means builders can obtain and use it nevertheless they please (with sure limitations). That’s in distinction to fashions like Anthropic’s Claude, Google’s Gemini, xAI’s Grok, and most of OpenAI’s ChatGPT fashions, which might solely be accessed through APIs.
Within the curiosity of giving builders alternative, nevertheless, Meta has additionally partnered with distributors, together with AWS, Google Cloud, and Microsoft Azure, to make cloud-hosted variations of Llama accessible. As well as, the corporate publishes instruments, libraries, and recipes in its Llama cookbook to assist builders fine-tune, consider, and adapt the fashions to their area. With newer generations like Llama 3 and Llama 4, these capabilities have expanded to incorporate native multimodal assist and broader cloud rollouts.
Right here’s all the things you should find out about Meta’s Llama, from its capabilities and editions to the place you need to use it. We’ll hold this publish up to date as Meta releases upgrades and introduces new dev instruments to assist the mannequin’s use.
What’s Llama?
Llama is a household of fashions — not only one. The most recent model is Llama 4; it was launched in April 2025 and consists of three fashions:
- Scout: 17 billion lively parameters, 109 billion whole parameters, and a context window of 10 million tokens.
- Maverick: 17 billion lively parameters, 400 billion whole parameters, and a context window of 1 million tokens.
- Behemoth: Not but launched however will have 288 billion lively parameters and a couple of trillion whole parameters.
(In information science, tokens are subdivided bits of uncooked information, just like the syllables “fan,” “tas” and “tic” within the phrase “fantastic.”)
A mannequin’s context, or context window, refers to enter information (e.g., textual content) that the mannequin considers earlier than producing output (e.g., extra textual content). Lengthy context can forestall fashions from “forgetting” the content material of latest docs and information, and from veering off subject and extrapolating wrongly. Nonetheless, longer context home windows also can outcome within the mannequin “forgetting” sure security guardrails and being extra susceptible to provide content material that’s in step with the dialog, which has led some customers towards delusional pondering.
For reference, the 10 million context window that Llama 4 Scout guarantees roughly equals the textual content of about 80 common novels. Llama 4 Maverick’s 1 million context window equals about eight novels.
Techcrunch occasion
San Francisco
|
October 27-29, 2025
All of the Llama 4 fashions had been educated on “large amounts of unlabeled text, image, and video data” to provide them “broad visual understanding,” in addition to on 200 languages, in keeping with Meta.
Llama 4 Scout and Maverick are Meta’s first open-weight natively multimodal fashions. They’re constructed utilizing a “mixture-of-experts” (MoE) structure, which reduces computational load and improves effectivity in coaching and inference. Scout, for instance, has 16 consultants, and Maverick has 128 consultants.
Llama 4 Behemoth consists of 16 consultants, and Meta is referring to it as a instructor for the smaller fashions.
Llama 4 builds on the Llama 3 collection, which included 3.1 and three.2 fashions extensively used for instruction-tuned purposes and cloud deployment.
What can Llama do?
Like different generative AI fashions, Llama can carry out a variety of various assistive duties, like coding and answering fundamental math questions, in addition to summarizing paperwork in at the very least 12 languages (Arabic, English, German, French, Hindi, Indonesian, Italian, Portuguese, Hindi, Spanish, Tagalog, Thai, and Vietnamese). Most text-based workloads — assume analyzing massive information like PDFs and spreadsheets — are inside its purview, and all Llama 4 fashions assist textual content, picture, and video enter.
Llama 4 Scout is designed for longer workflows and big information evaluation. Maverick is a generalist mannequin that’s higher at balancing reasoning energy and response pace, and is appropriate for coding, chatbots, and technical assistants. And Behemoth is designed for superior analysis, mannequin distillation, and STEM duties.
Llama fashions, together with Llama 3.1, could be configured to leverage third-party purposes, instruments, and APIs to carry out duties. They’re educated to make use of Courageous Seek for answering questions on latest occasions; the Wolfram Alpha API for math- and science-related queries; and a Python interpreter for validating code. Nonetheless, these instruments require correct configuration and aren’t routinely enabled out of the field.
The place can I take advantage of Llama?
If you’re trying to merely chat with Llama, it’s powering the Meta AI chatbot expertise on Fb Messenger, WhatsApp, Instagram, Oculus, and Meta.ai in 40 international locations. High-quality-tuned variations of Llama are utilized in Meta AI experiences in over 200 international locations and territories.
Llama 4 fashions Scout and Maverick can be found on Llama.com and Meta’s companions, together with the AI developer platform Hugging Face. Behemoth continues to be in coaching. Builders constructing with Llama can obtain, use, or fine-tune the mannequin throughout many of the well-liked cloud platforms. Meta claims it has greater than 25 companions internet hosting Llama, together with Nvidia, Databricks, Groq, Dell, and Snowflake. And whereas “selling access” to Meta’s overtly accessible fashions isn’t Meta’s enterprise mannequin, the corporate makes some cash via revenue-sharing agreements with mannequin hosts.
A few of these companions have constructed extra instruments and providers on prime of Llama, together with instruments that permit the fashions reference proprietary information and allow them to run at decrease latencies.
Importantly, the Llama license constrains how builders can deploy the mannequin: App builders with greater than 700 million month-to-month customers should request a particular license from Meta that the corporate will grant on its discretion.
In Might 2025, Meta launched a new program to incentivize startups to undertake its Llama fashions. Llama for Startups provides corporations assist from Meta’s Llama group and entry to potential funding.
Alongside Llama, Meta offers instruments meant to make the mannequin “safer” to make use of:
- Llama Guard, a moderation framework.
- CyberSecEval, a cybersecurity threat evaluation suite.
- Llama Firewall, a safety guardrail designed to allow constructing safe AI techniques.
- Code Defend, which offers assist for inference-time filtering of insecure code produced by LLMs.
Llama Guard tries to detect probably problematic content material both fed into — or generated — by a Llama mannequin, together with content material referring to legal exercise, baby exploitation, copyright violations, hate, self-harm and sexual abuse. That mentioned, it’s clearly not a silver bullet since Meta’s personal earlier tips allowed the chatbot to have interaction in sensual and romantic chats with minors, and a few stories present these turned into sexual conversations. Builders can customise the classes of blocked content material and apply the blocks to all of the languages Llama helps.
Like Llama Guard, Immediate Guard can block textual content meant for Llama, however solely textual content meant to “attack” the mannequin and get it to behave in undesirable methods. Meta claims that Llama Guard can defend in opposition to explicitly malicious prompts (i.e., jailbreaks that try to get round Llama’s built-in security filters) along with prompts that include “injected inputs.” The Llama Firewall works to detect and forestall dangers like immediate injection, insecure code, and dangerous device interactions. And Code Defend helps mitigate insecure code solutions and provides safe command execution for seven programming languages.
As for CyberSecEval, it’s much less a device than a set of benchmarks to measure mannequin safety. CyberSecEval can assess the chance a Llama mannequin poses (at the very least in keeping with Meta’s standards) to app builders and finish customers in areas like “automated social engineering” and “scaling offensive cyber operations.”
Llama’s limitations
Llama comes with sure dangers and limitations, like all generative AI fashions. For instance, whereas its most up-to-date mannequin has multimodal options, these are primarily restricted to the English language for now.
Zooming out, Meta used a dataset of pirated e-books and articles to coach its Llama fashions. A federal decide not too long ago sided with Meta in a copyright lawsuit introduced in opposition to the corporate by 13 ebook authors, ruling that using copyrighted works for coaching fell below “fair use.” Nonetheless, if Llama regurgitates a copyrighted snippet and somebody makes use of it in a product, they may probably be infringing on copyright and be liable.
Meta additionally controversially trains its AI on Instagram and Fb posts, photographs and captions, and makes it troublesome for customers to decide out.
Programming is one other space the place it’s smart to tread frivolously when utilizing Llama. That’s as a result of Llama may — maybe extra so than its generative AI counterparts — produce buggy or insecure code. On LiveCodeBench, a benchmark that assessments AI fashions on aggressive coding issues, Meta’s Llama 4 Maverick mannequin achieved a rating of 40%. That’s in comparison with 85% for OpenAI’s GPT-5 excessive and 83% for xAI’s Grok 4 Quick.
As at all times, it’s finest to have a human knowledgeable evaluate any AI-generated code earlier than incorporating it right into a service or software program.
Lastly, as with different AI fashions, Llama fashions are nonetheless responsible of producing plausible-sounding however false or deceptive info, whether or not that’s in coding, authorized steering, or emotional conversations with AI personas.
This was initially printed on September 8, 2024 and is up to date frequently with new info.
