Everything a working professional needs to actually use Meta AI and Llama — including the parts that changed while everyone else's course stayed the same.
Two separate product lines, one confusing name. Know which is which, what it costs, what the license really says, and who can read your conversation.
The assistant inside WhatsApp, Instagram, Messenger and Facebook. It runs on Muse Spark, the family Meta announced in April 2026 — closed weights you cannot download.
Models you download and run on your own hardware. Meta's position: unless you send inputs or outputs to Meta or publish them, Meta "cannot access, view, or use" them.
Meta AI no longer runs on Llama. The two lines answer the privacy question in opposite ways, which is why chapters six and seven treat them separately.
By the end of this chapter: you can find Meta AI on every app you already have, and you can explain the one change that makes most Meta AI training obsolete.
Meta AI is the assistant Meta puts inside apps you already open every day — WhatsApp, Instagram, Messenger, Facebook — plus a standalone app, a website, and their glasses. You don't sign up for anything: Meta says "no separate account is needed beyond your existing Meta account." That's the whole pitch, and it's a real one. It's already on the phone in your pocket.
Where it is, and how you actually start it
Same assistant, six front doors. The invocation differs in each one, which is the bit people fumble in front of a room.
meta.ai and sign in with your existing Meta account.In a group chat, you summon it by name. Type @Meta AI in a WhatsApp or Messenger group and it joins that message. Useful — and it comes with a catch big enough that it gets its own box below.
By the end of this chapter: you can give an honest answer to "is Meta AI free?" — which is no longer a one-word answer.
For years the answer was simply "yes." That changed. Meta now has paid subscription tiers and has begun capping the heavier features. Here's their own wording, because the nuance is the whole answer: "Meta AI is free to use for everyday use. We've started testing usage limits on some of our more compute intensive features. When people hit their daily / monthly limit, they'll have the option to purchase a subscription to unlock more usage before their free limit resets."
Asking questions, drafting, everyday image creation — no subscription, no credit card. For most people, most days, the free tier is the whole product.
The compute-intensive features — lots of image and video generation, and the deeper reasoning mode — now have daily and monthly limits. Hit one and you're offered an upgrade or a wait.
Personal tiers are Meta One Plus and Meta One Premium. There are separate business and creator tiers: Essential, Advanced, Expert and Max.
Meta's list: generate more images and videos with expanded access to advanced models, more advanced reasoning via Thinking mode, and a larger monthly AI usage allowance.
Meta describes Meta One as in "limited testing" and says it "isn't available everywhere yet." It's mobile only — iPhone or Android — and requires a personal, not business, Meta account. Don't promise it to a room that can't get it.
Straight from Meta: "Unused usage for AI benefits from a previous billing cycle does not roll over." Use it or lose it each cycle — so buy the tier that matches a normal month, not your busiest one.
By the end of this chapter: you've used all three modes on purpose, and you've held a real voice conversation instead of typing.
Meta AI's biggest advantage isn't raw capability — it's that it's already on the phone in your hand, and it's genuinely good at being talked to rather than typed at. That changes when you use it: in the car, on the shop floor, walking the yard.
Three modes — pick one on purpose
In the Meta AI app, tap the mode name in the bottom-right corner to switch. Most people never notice it's there and stay on the default forever.
The fast lane. Short questions, quick drafts, "what's the conversion for…" — you want the answer now, not a considered essay.
Meta's deeper reasoning mode, for complex questions and multi-step problems. It's also one of the features that counts against the paid allowance — so use it when the question earns it.
Tuned for products — comparing options, finding something specific. Meta AI can also pull in recommendations from Reels and maps as you talk.
Voice — where it genuinely pulls ahead
Meta's description of what Muse Spark unlocked: you can "interrupt, switch topics, or swap languages" mid-conversation. That's much closer to talking to a person than to operating a device.
Tap the microphone icon before sending, and talk. You don't have to wait for it to finish — interrupt it, the way you would a colleague who's going off track.
During a voice conversation, tap the camera icon to ask about your surroundings — a label, a part, a form, a room. This came to glasses first and reached the app in 2026.
Say "Hey Meta" or tap and hold the touchpad. Follow-up questions don't need the wake word again — so it's a conversation, not a series of commands.
There's an opt-in full-duplex demo under , app only. And a plain limitation worth remembering before you demo it: voice conversations can't generate images.
By the end of this chapter: you've made and edited a real image, and you know what's invisibly attached to every one you create.
Image generation is where Meta AI stopped being a novelty. It runs on Muse Image, and the interesting part is how it works — Meta's own description: it "doesn't just build an image; it thinks through your prompt first," planning the layout, looking up real-time web context, and blending multiple references before it draws anything.
Legible text inside the image — the classic AI weak spot — plus working QR codes, precise object removal, and blending several reference images into one composition.
You don't have to re-describe the whole picture. Sketch or mark on the image to show what you want changed, or work from the presets panel. Point at the problem instead of explaining it.
Redesign a room and the results can be made shoppable against Facebook Marketplace — the generated look connects to things you can actually buy. That's a real workflow, not a demo.
Meta's newsroom wording: free for everyday creation, and for people who want to create more, it's part of the subscription plans. Volume is exactly where the caps bite.
meta.ai/identification. Teach this as a feature, not a warning: it means an AI-generated image stays identifiable as one long after it leaves your hands. If you're putting AI images into customer-facing work, that's a fact you want to know before someone else discovers it.By the end of this chapter: you know what Meta actually offers a business, what's generally available versus waitlisted, and the one disclosure you owe your customers.
This is the chapter that matters if you sell anything. Meta's business play isn't a chatbot on your website — it's an agent living inside the messaging apps your customers already use to contact you. Meta says "more than one million businesses" are already running one on WhatsApp and Messenger.
Meta's list: answer questions specific to your business, make product recommendations from your catalog, book appointments, qualify incoming leads, let a human step in when you choose — and close sales.
Meta: businesses can activate an agent and "getting started is free." They also say that "in the coming months" it moves to paid subscriptions with options for businesses of every size. Set it up now, budget for later.
The customer-response agent is being expanded to businesses of all sizes globally — that's the part you can actually count on today.
The other half — morning briefings and running your operation for you — is limited to "a select number of businesses" with a waitlist. Don't build a plan on this half yet.
The two other business tools — and their real status
Meta ships a lot and announces earlier than it ships. Here's what's actually usable versus what's still a headline.
Connects your agent to hundreds of systems — Shopify, Zendesk, Shopee — so it can act on your behalf, with enterprise-grade controls.
Help with ads and account questions inside Business Suite, Ads Manager and Instagram Boost. Still beta, English-first, rolling out through 2026.
Meta says advertisers will tap Muse Image through Advantage+ creative "in the coming weeks." Announced, not shipped. Check before you promise it.
By the end of this chapter: you can explain the split between Meta AI and Llama, name the current models, and say honestly whether running one yourself makes sense for you.
Now the part that separates people who know this subject from people who read about it. Llama is not Meta AI. Llama is a family of models you can download and run on your own hardware — the reason Meta matters to businesses that can't send their data anywhere. Meta AI is the consumer assistant, and it runs on Muse. Two lines, one company, completely different jobs.
Muse Spark and Muse Image power the assistant. They are closed weights — available by API to select partners in private preview. Meta says they "hope to open-source future versions," which is a hope, not a commitment.
Llama models you can run on your own machines. Nothing leaves your building. This is the entire reason Llama exists in a business conversation.
Meta's own answer, verbatim: "Unless you provide inputs or outputs directly to Meta, or publish them, Meta cannot access, view, or use your inputs or outputs." For regulated or confidential work, that sentence is the whole pitch.
Heads-up if you've bookmarked anything: llama.com now redirects — the models and docs live at developer.meta.com/ai. Old links in old courses simply bounce.
The current lineup
Two generations are live. The numbers matter less than the shape: one family is built for scale, one for running small and cheap.
109B total parameters, and a class-leading 10 million token context window. Meta markets it as fitting single-GPU efficiency — the realistic starting point for a business.
400B total parameters — Meta's most powerful open multimodal model. Not a single-GPU proposition; this one needs real infrastructure.
The multilingual workhorse, and Meta's stated replacement for Llama 3.1 70B, which has been retired from the download page.
The lightweight pair, cheap enough to run on mobile and edge devices. A model on a handheld scanner is a real thing you can do.
Open multimodal models that reason over high-resolution images and return text. Photograph a damaged pallet; get a written condition note.
Llama 4 comes with Llama Guard 4 and Prompt Guard 2. Language support differs by generation: 3.1, 3.2 and 3.3 cover 8 languages; Llama 4 covers 12.
Getting one — and the honest hardware answer
Downloading is free and gated by a form. Running it well is where the real cost lives, and nobody should sell you past that.
developer.meta.com/ai. The form wants your legal name, date of birth, email, country, organization and job title — Meta warns that acronyms or missing corporate identifiers can block your access.If you want Muse instead of Llama — Meta offers the Meta Model API, currently in public preview for US developers only, and OpenAI-SDK compatible. Published rates: $1.25 per million input tokens, $0.15 cached input, $4.25 output, and $2.50 per 1,000 web-search-grounded queries, with $20 in free credits to start. Rates move — check before you budget.
By the end of this chapter: you know what the Llama license actually permits, the one clause that stops European companies cold, and what happens to everything you type into Meta AI.
People call Llama "open source." It isn't — and Meta doesn't claim it is. They call it "a bespoke commercial license." The good news: for a normal business, it's genuinely permissive. The catch is that the parts people repeat at conferences are usually the parts they get wrong.
You may use, reproduce, distribute, copy, create derivative works of, and modify the models — worldwide, royalty-free. And you own your derivatives: as between you and Meta, modifications you make are yours.
Everyone says you can't. That was true for Llama 2 and 3. From Llama 3.1 onward it's allowed, provided you include the correct attribution. If you heard otherwise, you heard about an old license.
Cited constantly as a blocker. Read it properly: it applies only if, on the model's release date, you had over 700 million monthly active users. It's a snapshot test aimed at Meta's rivals. For a normal business it is simply irrelevant.
You must prominently display "Built with Llama," ship a copy of the license and a NOTICE file, and if you fine-tune and distribute a model, its name must start with "Llama."
The clause that stops European companies
This is the single most business-relevant term in the whole license, and it's missing from almost every article written about it.
What you may not use it for
The Acceptable Use Policy is part of the license, so breaking it breaks your rights. Most of it is the obvious list — but three items catch ordinary businesses by surprise.
No unlicensed practice of a profession — Meta names financial, legal and medical/health specifically. An AI that answers customers' financial or medical questions is squarely in this territory.
Not for the operation of critical infrastructure, transportation technologies, or heavy machinery — and Meta is explicit that this holds "regardless of government certifications."
You may not represent that outputs are human-generated, and you must disclose known dangers of your AI system to end users. That's a marketing-copy rule as much as a legal one.
Illegal activity, harassment, unlawful discrimination in hiring/credit/housing, malware, weapons, disinformation, impersonation and fake reviews. If you need an exception, Meta reviews requests case by case.
What happens to what you type
Different question, same importance. And the answer differs sharply depending on whether you're using the assistant or running a model yourself.
Meta's own words: "When information is shared with AIs, the AIs will sometimes retain and use that information," and "Do not share information that you don't want the AIs to use and retain." Meta also reviews conversations, automatically or manually.
Meta says of it: "no one — not even Meta — can read your conversations," conversations aren't saved, and messages disappear by default. Note it's still rolling out across WhatsApp and the Meta AI app, so don't assume you have it.
Users there receive notice and an objection form under GDPR to object to their public data being used to train Meta's models, along with access, deletion and correction rights. Elsewhere, no equivalent training opt-out is documented.
Run it yourself and none of the above applies: unless you send inputs or outputs to Meta or publish them, Meta states it cannot access, view or use them. If the data can't leave, that's your answer — and it's the strongest argument for Llama in a regulated business.
Say the quiet part out loud. Outside the EU and UK, Meta documents no toggle that removes your conversations from training. Deletion and discipline are the controls you have — so the discipline is the real control.