Meta AI: Muse, Llama and Personal Superintelligence

Meta Is Building an AI That Knows You - and Acts for You

From Llama to Muse, AI glasses and personal superintelligence, Meta is trying to turn artificial intelligence into a layer that follows people across apps, devices and daily life.

Meta AI is no longer just a chatbot or a model family. In 2026, Meta is building a broader ecosystem that connects social platforms, smart glasses, personal AI agents, and large-scale infrastructure.

For years, Meta's AI story was easy to summarize. Facebook had a first-rate research lab, recommendation systems that shaped what billions of people saw, and an increasingly influential family of language models called Llama. Meta mattered in AI, but it never captured the public imagination the way ChatGPT did.

By 2026, that description no longer fits. Meta is not simply trying to build a better chatbot. It is building AI into WhatsApp and Instagram, teaching it to understand what a camera sees, remember preferences, use a browser, connect to calendars and email, and keep working after the user closes the app. Its newest consumer bet, Muse, is described as a personal AI agent rather than a conversational assistant.

So the most interesting question is no longer whether Meta can win another benchmark. It is whether the company can combine social networks, messaging, creators, ads, smart glasses and billions of daily users into an AI layer that feels unusually personal - and unusually hard to leave.

That could be genuinely useful. It could also make mistakes, privacy failures and unwanted automation much more consequential. Meta's next chapter is therefore less about winning an abstract AI race than about a practical test: how much of our digital lives are we willing to hand to software that knows enough about us to act on our behalf?

Meta stopped chasing a chatbot

Meta AI makes more sense if you stop treating it as a single product. It is really a stack with three layers.

Layer	What Meta is building	Why it matters
Models	Llama, Muse Spark, Muse Image, Muse Glimmer and other foundation models	The intelligence that reasons, sees, creates and uses tools.
Surfaces	Meta AI, WhatsApp, Instagram, Facebook, Messenger, Threads and AI glasses	Distribution: AI appears where people already communicate and create.
Agents	Meta AI task features and the separate Muse personal agent	The shift from answering questions to carrying out multi-step work.

 

The first layer is familiar: models. The second is Meta's unusual advantage: distribution. The third is newer - agents that do more than produce text. They can break a goal into steps, use tools, monitor progress and, in some cases, pause for permission before an irreversible action. That shift became central to Meta's product strategy in 2026.

From FAIR to Llama: how Meta got here

Meta's AI history did not begin with generative chatbots. Facebook created its FAIR research lab in 2013, long before the current AI boom. The lab became known for foundational work in computer vision, language, self-supervised learning, robotics and open research tools. PyTorch, originally developed inside Facebook, became one of the dominant frameworks for modern machine learning.

The turning point in Meta's public AI identity came in February 2023 with LLaMA. The family was designed to show that comparatively efficient models, trained on carefully selected data, could compete with much larger systems. Llama 2 followed later that year with broader commercial access. Llama 3 arrived in 2024, and Llama 4 in 2025 introduced natively multimodal mixture-of-experts models such as Scout and Maverick.

Llama mattered less because it powered a consumer chatbot and more because developers could actually build with it. Researchers, startups and companies could run the weights, fine-tune them and deploy systems outside Meta's own products. That made Llama one of the reference points of the open-model ecosystem.

If you want the broader foundations first, see Next Horizon's Understanding Artificial Intelligence, which explains machine learning, foundation models and generative AI in simpler terms.

The open-model story is more complicated than the slogan

Meta has long presented openness as a defining part of its AI strategy, and that approach helped Llama spread. The terminology, however, needs care. Llama models have generally been released as open weights under Meta's own community licenses, not as unrestricted open-source software in the strict sense used by organizations such as the Open Source Initiative. Those licenses include conditions that can fall short of the formal Open Source Definition.

The difference is more than semantics. Open weights usually means developers can download the model and build on it. Open source traditionally carries broader freedoms to use, modify and redistribute the technology. Meta has been far more open than an API-only provider, but 'Llama is completely open source' is still too simple a description.

Then the center of gravity moved again. Llama 4 remained important, but Meta's most aggressive consumer and agent work shifted in 2026 toward a new family from Meta Superintelligence Labs: Muse. Llama did not disappear; it simply stopped being the whole story.

Meta’s AI strategy has evolved from research-led work at FAIR to open-weight Llama models, standalone consumer AI products, and the 2026 push toward Muse and personal AI agents.

The 2025 pivot: from "open AI" to personal superintelligence

Meta's AI strategy became much more aggressive in 2025. The company invested $14.3 billion in Scale AI and brought its CEO, Alexandr Wang, into Meta to lead a new superintelligence effort. Meta Superintelligence Labs was then organized around a broader goal: not merely answering more questions, but building what Mark Zuckerberg calls personal superintelligence.

The phrase sounds grand. The product idea is easier to grasp. Meta wants an assistant that can help one person pursue long-running goals - learning, creating, organizing life, communicating, shopping and eventually acting through software and devices - rather than one giant system that tries to automate everything for everyone.

That ambition plays directly to Meta's strongest asset. OpenAI, Anthropic and Google can all build excellent models. Meta already owns many of the places where people's digital social lives happen: WhatsApp, Instagram, Facebook, Marketplace, Threads and a growing set of wearable devices. With permission, that context could make an assistant feel less like a blank chatbot and more like software that already understands what matters to its user.

Muse Spark: Meta rebuilt the model stack around agents

The first model from Meta Superintelligence Labs, Muse Spark, arrived in April 2026. Meta described it as a relatively small, fast model built to test a new scaling approach while still handling reasoning and multimodal tasks. The more revealing detail was where it was aimed: directly at Meta's products, not mainly at research demos.

The releases then came quickly. Muse Spark 1.1 in July emphasized tool use, computer use, coding and multimodal reasoning. Version 1.2 in August pushed further into long-running coding and visual reasoning. Muse Spark 1.3, released on September 2, focused on longer agent workflows: tracking several goals, using tools across messy information, noticing when a plan is going wrong and asking before consequential actions.

Meta also released Muse Glimmer, a 30-billion-parameter agentic model designed to run locally on a consumer computer, while promising larger models and future open-weight releases. The pattern suggests a split architecture: large cloud models for the hardest work, smaller local agents where speed, cost, privacy or offline access matter more.

The deeper question - what these models actually represent internally when they reason and use tools - connects directly to Next Horizon's Inside the Black Box: How Large Language Models "Think".

Meta AI is becoming a layer across the company

Meta AI has changed almost as quickly as the models beneath it. The standalone app launched in April 2025 on Llama 4 with voice conversation, web search, image generation, memory and personalization. Meta said the assistant could remember facts a user chose to share and, in supported markets, use information from connected Facebook and Instagram activity to make answers more relevant.

By 2026, Muse Spark had replaced Llama 4 at the center of the consumer experience. Meta AI added Instant and Thinking modes, better visual understanding, research-style tasks, shopping tools and parallel subagents. In July, Meta added agentic features that could connect to email and calendars, prepare briefings, research a topic, generate slides and keep recurring tasks running without a fresh prompt each time.

This is the part of Meta's strategy that a benchmark cannot capture. The assistant is being threaded through search bars, group chats, social posts, Marketplace, WhatsApp, Instagram, Facebook, Messenger, Threads and AI glasses. Meta does not need users to form a completely new habit; it can insert AI into habits they already have.

That reach is a formidable advantage, but it also raises the stakes. More context can make an assistant more useful. It can also make a mistaken action, a security flaw or a misunderstood permission far more damaging.

Meta wants AI to become a constant layer across daily life — from chat and social discovery to shopping, scheduling, smart glasses, and creative tools.

Muse changes the product category: the assistant can now act

On September 8, 2026, Meta launched Muse in the United States. It is not simply Meta AI with a new name. Muse is a dedicated personal agent with its own app, WhatsApp access, a browser and a secure virtual computer in the cloud. The user gives it a goal rather than a sequence of clicks.

Ask it to organize a trip and it can search, compare options, fill forms and return when approval is needed. Ask it to sell a car and it can help prepare the listing and negotiate. Ask it to lower a bill and it can work through the process in the background. Meta says Muse can continue after the app is closed and return when something changes or when a sensitive step - sending an email or making a purchase, for example - needs confirmation.

This is the real jump from chatbot to agent. A chatbot turns information into an answer. An agent turns an objective into actions. The difference becomes much less abstract once money, credentials, calendars or personal relationships enter the workflow. A slightly wrong answer may be irritating; a slightly wrong autonomous action can be costly.

The security architecture is part of the product, not an afterthought

Meta built Muse around what it calls Muse Secure VM: a dedicated cloud virtual machine that contains the agent's working environment and data from connected services. A separate Sentinel agent is meant to approve outbound actions and trigger permission checks when needed. Meta says Muse cannot directly see passwords or payment credentials, provides an audit trail, lets users revoke app access and keeps Muse conversations and VM data separate from advertising systems.

Those are sensible safeguards, not a guarantee. Reuters reported that Meta delayed Muse earlier in 2026 to strengthen security, and that internal testers still found reliability and privacy failures close to launch, including unexpected behavior around sensitive personal data. Meta's position was that the product had crossed a minimum safety threshold, not that errors had been eliminated.

That is the standard agents should be judged against. Safeguards matter only if they survive thousands of messy real-world interactions: websites with prompt injections, changing interfaces, confusing permission flows and models that sometimes believe a task is finished when it is not.

The next step for Meta AI is not just answering questions but taking action. This concept shows how a personal AI agent could plan, research, organize tasks, and interact with outside services while keeping the user in control.

AI glasses may be Meta's most important hardware bet

If a personal AI is supposed to understand a person's world, a phone is a surprisingly limited sensor. It spends much of the day in a pocket. Glasses can see roughly what the wearer sees, hear what the wearer hears and respond without forcing someone to stop and type. That helps explain why Meta increasingly treats glasses as a natural home for an always-available assistant.

The hardware has moved well beyond the original Ray-Ban Meta camera glasses. Ray-Ban Meta Gen 2 improved battery life and video. Meta Ray-Ban Display added an in-lens display paired with an EMG wristband that reads subtle muscle signals for control. In 2026 Meta broadened the lineup with prescription-focused models, Oakley performance glasses and other Meta Glasses products, while Muse Spark improved visual understanding.

Put those pieces together and ambient computing starts to look plausible. Instead of reaching for a phone to identify an object, translate a conversation, read a reminder or ask for directions, the assistant can respond in the moment. Meta already offers live translation, visual questions and accessibility features, and says Muse is coming to AI glasses as well.

The same hardware creates a social problem that phones mostly avoid: bystanders may not know when a camera is active or what an AI system is processing. Meta uses a capture LED and says covering it disables the camera, but the broader question remains. As glasses become more useful, expectations around consent and privacy in workplaces, public spaces and intimate settings will have to catch up.

Meta’s long-term vision goes beyond apps on a phone. AI glasses could make translation, navigation, and contextual assistance part of the real world around us.

Meta is building a media-generation stack too

Personal AI is only one branch of Meta's strategy. Meta Superintelligence Labs also launched Muse Image in July 2026 and previewed Muse Video, a video model with native audio. Meta describes Muse Image as agentic because it can use tools, search and self-refinement instead of mapping a prompt directly to pixels in one pass.

The strategic value is obvious: Meta already owns enormous creation and distribution surfaces. If an image model lives inside Instagram Stories or WhatsApp, users do not need to generate an asset elsewhere and then move it back into Meta's ecosystem. The same logic can extend to video, ads, creator tools and personalized media.

Meta One, launched on September 15, makes the business model clearer. The core Meta apps and Meta AI remain free, while paid plans raise AI limits and add creator or business features. That points to a hybrid model: AI can keep improving engagement and advertising, while the heaviest creation and agent features become subscription products.

Meta's real advantage may be context, not raw intelligence

It is tempting to reduce the AI race to a leaderboard: who writes the best code, tops a benchmark or answers the hardest science question. Meta may be playing a different game. A model that is slightly weaker in isolation can still be more useful if it knows who you talk to, which creators you follow, what you are shopping for, what is on your calendar and which places matter to you.

Meta has leaned openly into that advantage. The Meta AI app can use remembered preferences and, where enabled, information users have already shared across Meta products. Interactions with Meta's AI can also influence content and ad recommendations in some markets. In Europe, Meta says it uses public adult content and interactions with Meta AI for model training while providing mechanisms to object under applicable rules.

This is Meta's moat and its pressure point at the same time. More context can make the assistant better, but every new source raises another question. Did the user expect that information to shape an AI response? Can they see what the system remembers? Can they delete it? Could something said casually in a chat later affect recommendations or advertising? And what happens when an agent combines information that used to live in separate silos?

Meta has added controls and makes stronger promises for Muse: users can choose connections, revoke access, opt out of training and keep Muse VM data away from ad systems. Whether personal AI earns lasting trust will depend less on the existence of those controls than on whether ordinary users can understand and use them.

The infrastructure race behind the friendly assistant

The visible side of Meta AI is a chat window or a pair of glasses. Behind it sits an industrial system of data centers, accelerators, memory, networking, cooling and electricity. Meta's spending shows how central that infrastructure has become to the strategy.

Reuters reported that Meta expected to spend as much as $145 billion on AI infrastructure in 2026 and was targeting roughly seven gigawatts of computing capacity during the year, with plans to double total capacity again in 2027. Meta is also pushing its custom AI-chip program forward to reduce dependence on outside suppliers and lower the cost of serving AI across products used at enormous scale.

This is why AI strategy increasingly looks like energy strategy too. Training a frontier model is only one expense. Serving personalized AI continuously - voice, images, agents, glasses and background tasks - may require even more inference capacity over time. Consumer AI will be shaped not only by software quality, but by who can afford to run that software for billions of interactions.

What Meta still has to prove

Meta now has almost every ingredient an AI company could ask for: frontier researchers, open-model credibility, massive distribution, an advertising business, consumer hardware, an agent platform and one of the world's largest computing budgets. None of that guarantees a good product.

·         Reliability: agents must complete long tasks without silently failing, losing constraints or inventing success.

·         Security: browser-using agents create new attack surfaces, especially around prompt injection, credentials and payments.

·         Trust: Meta is asking users to give an AI system access to some of the most personal parts of digital life while the company still carries a long history of privacy controversy.

·         Model leadership: Meta's Llama 4 rollout was less dominant than the company hoped, and the new Muse family must prove it can compete consistently, not just in internal evaluations.

·         Openness: Meta must decide how much of its future frontier stack it is genuinely willing to release as open weights while simultaneously monetizing APIs, subscriptions and consumer products.

·         Economics: AI agents are expensive to run. The company must show that subscriptions, advertising gains and new hardware can justify infrastructure spending at this scale.

There is a useful warning inside Meta itself. Reuters reported in August 2026 that an ambitious internal effort to reorganize teams around AI agents ran into reliability, productivity and employee-trust problems. That does not make agents a dead end. It does show how much harder it is to turn an impressive demo into dependable infrastructure people use every day.

Meta vs ChatGPT, Gemini and Copilot: a different battlefield

Meta's closest rivals are not all fighting for exactly the same territory. OpenAI is building a broad AI platform around ChatGPT, tools and agents. Google can connect Gemini to Search, Android and Workspace. Microsoft can put Copilot inside enterprise software. Meta's clearest advantage lies in social context and wearable computing.

That makes simple model rankings less useful than they appear. The strategic question is not whether Muse Spark writes slightly better code than another model. It is whether an agent inside WhatsApp, Instagram and smart glasses can become more useful to ordinary people than an assistant that mainly lives in a separate app.

For another example of an AI ecosystem built around existing productivity software, see Next Horizon's Microsoft Copilot review.

What happens next?

The next stage is easy to imagine and much harder to predict. The safest way to think about it is to separate what Meta has actually announced from what remains speculation.

Over the next year

Meta has already said that larger Muse models are coming, Muse Spark open weights are on the roadmap, and Muse Video will expand beyond preview. Muse is expected to reach AI glasses. Meta also plans a more strongly encrypted Confidential VM in which, it says, even Meta will not be able to access the user's agent data. Meta One gives the company a ready-made subscription layer for higher AI usage.

Over the next few years

The likelier near-term future is not one dramatic leap to AGI, but a gradual blurring of the lines between assistant, browser, social network and operating system. A Meta agent could move from recommending a restaurant to coordinating friends, checking calendars, making a booking and guiding the group through glasses. A creator could move from an idea to images, video, publishing and audience analysis without leaving Meta's ecosystem. A small business could run customer-service, advertising and sales agents on the same underlying stack.

The longer-term bet

Meta's stated destination is personal superintelligence: an AI that understands an individual's context well enough to help with goals over months or years rather than minutes. Whether that will ever deserve the word 'superintelligence' is unknowable today. The product direction, however, is already visible: persistent memory, multimodal perception, long-running agents, physical-world interfaces and proactive assistance.

The harder question is not whether Meta reaches human-level or superhuman intelligence first. It is whether people decide that the convenience of an AI that knows their world is worth the access such an AI requires.

Conclusion: Meta is trying to own the layer between you and the internet

The old Meta AI story centered on research and Llama. The new one is broader: models, agents, media generators, developer APIs, custom chips, social integrations and wearable hardware, all built around the idea that AI should not wait inside a chatbot window. Meta wants it present wherever people already communicate, create, shop, navigate and make decisions.

Muse matters because it turns that strategy into a product. The assistant is no longer only something that answers. It can observe context, remember goals and carry out work across the internet.

If Meta succeeds, its biggest advantage may not be having the smartest model at every moment. It may be having AI in places billions of people already use - and increasingly on devices they wear. The same closeness is also the risk. The more a personal AI knows and the more it can do, the less room there is to treat security, errors and trust as secondary concerns.

Meta has now placed a serious bet on personal AI: a new generation of agentic models, enormous distribution and a product that can already act on a user's behalf. The next phase will be decided less by benchmark charts than by something harder to measure - whether people find that kind of AI useful enough, and trustworthy enough, to keep close to their everyday lives.

FAQ

What is Meta AI in 2026?

Meta AI is Meta's consumer AI assistant across meta.ai, the Meta AI app and parts of WhatsApp, Instagram, Facebook, Messenger, Threads and AI glasses. Its current experience is built around Meta's Muse model family rather than only Llama.

What is Muse?

Muse is Meta's separate personal AI agent launched in September 2026. It can use a browser and connected apps to perform multi-step tasks, work in the background and ask for approval before sensitive actions.

Is Llama still important to Meta?

Yes. Llama remains a major open-weight model ecosystem and is widely used by developers. But Meta's latest consumer and agentic product strategy has shifted toward models from Meta Superintelligence Labs, including Muse Spark and Muse Glimmer.

Are Meta's Llama models open source?

They are widely available as open-weight models under Meta licenses, but organizations such as the Open Source Initiative argue that those licenses do not meet the formal Open Source Definition. "Open-weight" is the more precise term.

Why are AI glasses important to Meta?

Glasses can give an assistant hands-free access to voice, visual context, translation and navigation. Meta sees them as a possible everyday interface for personal AI, although camera-equipped glasses also create important privacy questions for people nearby.