AI at Meta

What is AI at Meta?

AI at Meta is the official portal at ai.meta.com where Meta publishes everything it does in artificial intelligence. It is worth being precise about the category, because it is easy to mistake: this is not a single tool you sign up for and start using. It is a hub with three distinct kinds of content behind it — consumer products, frontier models, and open research — each of which lives somewhere else once you click through. The navigation tells you what kind of site this is: Products, AI Research, Resources, About and AI Developers.

The site's own description is precise about what it offers: use Meta AI to answer questions, create images, and complete tasks for free, plus AI research, models, and tools for developers. Read that sentence carefully and you get the whole structure. There is a consumer assistant, free to use. There is a body of research and a set of models. And there is a developer track. The portal is the index; the actual using happens in the Meta AI app, on meta.ai, inside WhatsApp or Instagram, or in a repository you clone.

Underneath sits a stated philosophy. Meta frames its goal as bringing personal superintelligence to everyone, putting that power in people's hands to direct toward what they value in their own lives — a position the company explicitly contrasts with those in the industry who believe superintelligence should be directed centrally toward automating all valuable work. Whatever you make of that as a claim about the future, it is the organising idea the portal is built around, and it explains why consumer products and fundamental research sit side by side on the same domain.

The institutional self-description is broad in scope. About page copy frames the mission as advancing AI for a connected world, spanning AI infrastructure, generative AI, natural language processing and computer vision, and stresses sharing research and engaging with the wider AI community rather than working alone. In practice that means a visitor arriving with a narrow question — which model should I use, is the assistant free, can I download the weights — has to navigate a site built to serve researchers, developers and ordinary phone users at the same time.

Core Features

• The Meta AI assistant: The consumer front door. It offers Instant mode for fast answers and Thinking mode for harder questions requiring advanced reasoning, plus a conversational voice mode supporting 13 languages and live visual answers about whatever the camera sees. You can upload several files and images at once — a lease, a spreadsheet, product photos to compare.
• The Muse model family: The current frontier line. Muse Spark 1.2 brings a terminal coding agent, Muse Code, that can plan, implement and validate complex, multi-file changes across large repositories. Muse Spark 1.1 is described as a multimodal reasoning model built for agentic tasks with major gains in tool and computer use.
• Multimodal reasoning at the core: Meta describes Muse Spark as a natively multimodal reasoning model with support for tool use, visual chain of thought and multi-agent orchestration. A Contemplating mode orchestrates multiple agents reasoning in parallel, and the health-related reasoning was developed with input from more than a thousand physicians.
• Image and video generation: Muse Image is presented as Meta's most advanced image generation model yet — following instructions faithfully, editing with precision, and composing from multiple references — alongside Muse Video and a media generation section of the site.
• Answers grounded in social content: Meta AI can cite public posts from Instagram, Facebook and Threads, grounding answers in what people are actually discussing right now. This is the genuine structural difference from assistants without a social graph behind them.
• Open research output: The research shelf is genuinely deep: Segment Anything 2 for video and image segmentation, DINOv3 for self-supervised vision backbones, V-JEPA 2 as a world model trained on video, and Movie Gen for media generation. Publications and the facebookresearch GitHub are linked from the footer.
• A developer track: A dedicated AI Developers destination sits in the top navigation, and the Resources section collects a blog, a Learning Hub and interactive demos for people who want to build rather than browse.

Use Cases

1. Working out which Meta model to use: The most common reason to land here. Anyone comparing Meta AI models starts at the homepage, which lists the current Muse releases with one-line descriptions and dates, so you can see at a glance what shipped recently and whether a coding agent, an image model or a general reasoning model fits your problem.
2. Getting the assistant onto your devices: The portal routes to the App Store, Google Play, a Mac desktop beta and the web app, and explains which Meta apps already have the assistant embedded.
3. Reading primary sources on a model: Rather than relying on secondary summaries, you can go straight to Meta's own launch posts, which state what a model does and — just as usefully — what the company does not claim about it.
4. Finding research artefacts to build on: For computer vision and multimodal work in particular, SAM 2, DINOv3 and V-JEPA 2 are downloadable research outputs with papers behind them. This is where Meta AI research is most directly reusable, reachable through the research and publications sections.
5. Understanding Meta's stated AI policy positions: The About and open source pages set out the responsible-AI principles and the open source argument, which matters if you are assessing the company as a supplier or writing about its strategy.
6. Tracking the company's direction: Because the portal collects launches, research and vision statements in one place, it is a reasonable way to follow where a major AI lab is heading over months rather than days.

How to use AI at Meta

1. Start by deciding which of the three tracks you want: Product, research or developer. The site is large, and the fastest route is knowing whether you want to use something, read something or build something.
2. To try the assistant, use the Try Meta AI button: It appears in the navigation and repeatedly through the homepage, and it hands off to the Meta AI app or meta.ai rather than running the assistant inside the portal itself.
3. Install where you actually work: Get the app from the App Store or Google Play, download the Mac beta if you want desktop dictation and window-aware help, or simply open the assistant inside WhatsApp, Instagram, Messenger or Facebook where it is already present.
4. For models, read the launch post before the summary: Each Muse release has an official blog entry describing capabilities and availability. These are the load-bearing documents; third-party write-ups vary in accuracy.
5. For research, go through AI Research and Publications: Projects link out to papers, model cards and code, with the facebookresearch GitHub organisation as the main code destination.
6. For building, follow the AI Developers link: It leaves ai.meta.com for Meta's developer site, which is where API access and integration documentation live rather than on the portal.
7. Check the date on everything: This portal moves quickly — the homepage news items span only a few weeks — so confirm that whatever you are reading still describes the current release.

Tips & Best Practices

• Do not treat the portal as the product: Nothing on ai.meta.com is the assistant itself. If you are evaluating capability, install the app or open meta.ai; the portal shows you marketing copy, not behaviour.
• Separate the model name from the product name: Meta AI is the assistant; Muse is the model family powering it. Conflating them makes comparisons with other vendors confusing, because a competitor's assistant and its model usually carry different names too.
• Prefer official launch posts to aggregator articles: When checking facts about a Meta model, the company's own blog is the primary source. Several widely-circulated third-party summaries assert benchmark numbers, context lengths and API prices that Meta's own posts do not state.
• Note what the company declines to say: The Muse Spark launch post specifies availability and capabilities but does not publish pricing or state whether weights are open. Silence on a point is information — it means the answer is not settled publicly, not that you should infer one.
• Use the research section if you need something you can actually run: The consumer models are services; the research releases are artefacts with papers and code. For reproducible work, the research shelf is the useful half of the site.
• Check regional availability before promising a feature: Assistant features roll out unevenly by country and by surface, and the portal's copy describes the fullest version of the product rather than what is live everywhere.
• Watch the open source page for drift: It remains the clearest statement of Meta's openness commitment, but read it against the current model line-up rather than assuming the policy covers every new release.

Who is AI at Meta for?

• Developers evaluating models: The primary audience for the technical half — people deciding whether a Muse model or an open research artefact fits a build, and who need capability descriptions from the source.
• AI researchers: Segment Anything 2, DINOv3, V-JEPA 2 and Movie Gen are the kind of releases that get cited and extended, and the publications index is aimed squarely at this group.
• Everyday users of the assistant: People who arrived wanting to know what Meta AI can do, whether it costs anything, and how to get it on a phone. For them the portal is a signpost to the app.
• Technology journalists and analysts: The portal collects launches, dates and official framing in one place, which makes it a reference point when reporting on the company's strategy.
• Businesses assessing Meta as an AI supplier: The responsible-AI principles, the open source position and the developer track together give a picture of how the company wants to be evaluated as a platform.
• Students and the AI-curious: The Learning Hub and demos exist for people who want to understand the field rather than ship something, though they are a small part of a large site.

Platforms

The portal itself is an ordinary website that works in any browser, but the products it points to have real platform detail worth knowing. Meta AI is available on iOS and Android, on the web, on desktop, on AI glasses, and inside Facebook, Messenger, Instagram and WhatsApp. On the App Store the assistant is published by Meta Platforms, Inc. in the Productivity category, rated about 4.7 stars from more than 212,000 ratings, and it requires iOS 17.2 or later. Its interface ships in 39 languages, and the current build was updated in August 2026. On Google Play the same assistant sits in Productivity with a 4.6-star rating from roughly 2.25 million reviews and more than 100 million downloads, carrying an in-app purchases label. A Mac desktop client exists in beta, capable of pointing at any window on screen to work through what you are doing or taking dictation into any app. On glasses, Meta states the experience is now powered by Muse Spark and covers hands-free questions, real-time translation and describing your surroundings, with a partnership with EssilorLuxottica behind the hardware. One footnote on store history: the App Store listing dates back to September 2021, well before the current assistant took its present shape — that is store metadata, not a statement about today's feature set.

Pricing & Plans

Meta states plainly on the portal that you can use Meta AI to answer questions, create images and complete tasks for free, and the app stores corroborate the entry point: the App Store lists the price as Free, and Google Play shows more than 100 million downloads at no cost to install. That said, Google Play also carries an in-app purchases label, so free-to-install does not mean nothing inside is ever paid. Meta does not publish a consumer subscription table on this portal — there is no pricing page listing tiers, quotas or monthly rates, and the FAQ entries on the assistant page that would address cost directly are loaded on interaction and could not be retrieved. On the developer side, the picture is deliberately incomplete too: at launch Muse Spark was available at meta.ai and in the Meta AI app, with an API offered only as a private preview to selected users, and the official post published no rates. Widely-shared third-party articles do quote specific per-token API prices, but neither Meta's own blog nor the authoritative press coverage confirms them, so they are not repeated here. The honest summary is this: the assistant is free to start using, some in-app purchases exist, and formal API commercial terms are not documented on this portal. Check the official site and Meta's developer destination for current terms before budgeting around any of it.

Alternatives

• OpenAI's portal and ChatGPT: The closest structural comparison — a company site that fronts both a consumer assistant and a model line-up. OpenAI publishes far more explicit pricing and API documentation; Meta's advantage is distribution through apps billions of people already have installed.
• Google AI and Gemini: Also spans research lab, model family and consumer assistant, with a similarly deep publication record. Google documents model tiers and developer pricing more openly, while Meta leans harder on being embedded in social apps and on glasses.
• Anthropic: A narrower, more enterprise-facing proposition with detailed model and safety documentation. It has no consumer social surface, which makes it the cleaner choice for teams who want a documented API and not an ecosystem.
• Hugging Face: The better destination if what you actually want is downloadable open models and community artefacts. Meta contributes to that world through its research releases, but Hugging Face is the neutral index across all vendors.
• Mistral AI and other open-weight labs: For teams whose requirement is genuinely open weights they can self-host, dedicated open-weight vendors are more predictable than a portal where the flagship line's licensing position is not stated.

Limitations & Considerations

• It is a portal, not a product: The single most common misunderstanding. You cannot evaluate the assistant here, and much of what looks like a feature list is a signpost to a different app or a different domain.
• Key commercial facts are undocumented: There is no consumer pricing table and no published API rate card on this site. The assistant FAQ that would answer the cost question directly does not render its answers into the page, so even a careful reader cannot resolve it from the portal alone.
• Marketing copy, not measured behaviour: Descriptions such as "our most advanced image generation model yet" are the company's own framing. There are no independent benchmarks on the portal, and capability claims should be tested against your own use case before you commit.
• The open source position is now split: Meta's open source page still carries Mark Zuckerberg's argument that open source is necessary for a positive AI future, so that power is not concentrated in the hands of a small number of companies. But that page's open line is Llama, while the current flagship Muse releases carry no stated position on open weights. The commitment and the frontier line no longer clearly coincide, and anyone choosing Meta for openness reasons should verify per model rather than trusting the policy page.
• Governance sits at platform level: The portal's Privacy Policy, Terms and Cookies links all point back to Meta's platform-wide policy centre rather than to standalone terms for this site. Your data relationship is with Meta as a whole, which is a different proposition from a standalone AI vendor.
• Real friction in the field: A visible minority of Play Store reviewers report heavy battery drain while the app runs in the background, along with the phone getting noticeably warm. Store ratings are high overall, but the complaint pattern is consistent enough to be worth knowing before installing on an older device.
• Privacy trade-offs are substantial: The Play Store data safety disclosure states the app may collect location, personal information and twelve other data categories, that data is encrypted in transit, and that deletion can be requested. Meta does present Incognito Chats as a private space, but that is one mode inside a broader data relationship, not the default posture.
• Fast-moving content ages quickly: Homepage news items span weeks, not years. Anything you read here — including this description — should be checked against the live site if the detail matters.
• Regional and surface variation: Features described in one place may not be available in your country or on your particular device, and the portal generally describes the fullest configuration rather than the typical one.

FAQ

Q1. Is AI at Meta a tool I can sign up for?

No, and this is the most important thing to understand about it. AI at Meta is Meta's official AI portal — an index of the company's products, models and research. There is no account to create here. When you click Try Meta AI you are handed off to the Meta AI app or meta.ai, and when you follow a research link you end up at a paper or a code repository.

Q2. Is Meta AI free to use?

The portal states directly that you can use Meta AI to answer questions, create images and complete tasks for free, and both app stores list the assistant as a free download. However, Google Play also shows an in-app purchases label, so some paid items exist inside the app. Meta does not publish a consumer pricing table on this portal, so the precise boundary between free and paid is not documented here.

Q3. What is Muse, and how does it relate to Meta AI?

Muse is the current family of Meta models; Meta AI is the assistant product that people actually use. Muse Spark is the reasoning model, Muse Image and Muse Video cover media generation, and Muse Code is a terminal coding agent introduced with Muse Spark 1.2. In short, Muse is the engine and Meta AI is the car.

Q4. What happened to Llama?

Llama remains the open source line and still anchors Meta's open source page, including the Llama Impact Program. The frontier work, however, has moved to the Muse family out of Meta Superintelligence Labs. Independent reporting places Muse Spark as the inaugural model from Meta Superintelligence Labs, the group led by former Scale AI chief executive Alexandr Wang, and connects the lab's creation to dissatisfaction with Llama's competitive progress.

Q5. Are Meta's current models open weights?

Not answerable from this portal for the Muse line. The official Muse Spark launch post describes capabilities and availability but says nothing about open weights, and the open source page's examples remain Llama and the FAIR research releases. If open weights are a requirement for your project, verify the licence for the specific model rather than relying on Meta's general open source stance.

Q6. Where can I actually use the Meta AI assistant?

On iOS and Android, on the web, on a Mac desktop beta, on Meta's AI glasses, and inside Facebook, Messenger, Instagram and WhatsApp. The glasses experience is described as powered by Muse Spark and covers hands-free questions, real-time translation and describing what is around you.

Q7. Is there an API for Meta's models?

At the Muse Spark launch, Meta said the model was available at meta.ai and in the Meta AI app, with an API opened as a private preview to selected users. The official post did not publish rates. Third-party articles circulate specific per-token prices, but those are not confirmed by Meta's own materials or by the authoritative press coverage, so treat them as unverified.

Q8. What research does Meta publish here?

A substantial body. The About page highlights Segment Anything 2 for segmenting objects in video and images, DINOv3 for self-supervised vision backbones, V-JEPA 2 as a world model trained on video, and Movie Gen for media foundation models. Publications and the facebookresearch GitHub organisation are linked from the site footer.

Q9. How does Meta handle privacy for AI conversations?

Meta offers Incognito Chats, described as a private space whose messages are processed in a secure environment that no one, including Meta, can access. Outside that mode, the Play Store data safety section states the app may collect location, personal information and twelve further categories, with encryption in transit and the ability to request deletion. Note that the portal's legal links go to Meta's platform-wide policies rather than site-specific terms.

Q10. How does Meta AI differ from ChatGPT or Gemini?

Two differences are structural rather than promotional. First, distribution: Meta AI is already inside WhatsApp, Instagram, Messenger and Facebook, so for many people there is nothing to install. Second, grounding: Meta AI can cite public posts from Instagram, Facebook and Threads, which no competitor can do. Against that, rivals publish clearer pricing and API documentation, and for developers who need documented commercial terms today, that gap matters.