Meta AI is the umbrella name for Meta’s artificial intelligence efforts, but under that name sit two genuinely different model families: Llama, Meta’s open-weight line running since 2023, and Muse, a new proprietary family that Meta Superintelligence Labs introduced in 2026. They differ in who can use them, how, and at what cost.

Llama: Meta’s Open-Weight Line

Llama (“Large Language Model Meta AI”) launched in February 2023 as a research-access model and has since gone through four major generations. From Llama 2 onward, Meta Platforms released the model weights for anyone to download, fine-tune, and run on their own hardware, subject to an acceptable-use license — not the same as a true open-source license (the Open Source Initiative has objected to Meta calling it “open source”), but far more open than a closed API. Llama 4, released in 2025, added a mixture-of-experts architecture and native multimodal (text-and-image) understanding. Because the weights are downloadable, Llama has become a default base for startups, researchers, and governments that want to run a capable model on their own infrastructure rather than send data to an external API.

Muse: Meta’s New Proprietary Line

Muse is different by design. Introduced in April 2026 by Meta Superintelligence Labs — the unit Meta built around AI researcher Alexandr Wang after a multibillion-dollar investment — Muse Spark was positioned as a natively multimodal reasoning model with tool use and a “Contemplating mode” that runs multiple agents in parallel. Meta says it reaches roughly the same capability as Llama 4 Maverick using an order of magnitude less compute. Unlike Llama, Muse Spark’s weights are not published; it runs only through Meta’s own infrastructure. A second model, Muse Image, generates pictures and ships free inside the Meta AI app, WhatsApp, and Instagram. A video-generation model, Muse Video, is planned.

Why Meta Is Running Two Tracks at Once

Keeping Llama open while building Muse closed lets Meta hedge its bets. Open weights turned Llama into a widely adopted developer ecosystem and a reference point in AI policy debates about who should control frontier models. But an open model is hard to monetize directly, and Meta’s leadership has acknowledged that Llama’s pace of improvement lagged rivals like OpenAI and Google. Muse is Meta’s answer: a closed, continuously updated model line it can charge for and tune specifically for the agentic and coding workloads that are currently the most competitive corner of the AI market.

How to Try Meta AI

The consumer product is free: Meta AI, running on Muse Spark, is built into Facebook, Instagram, WhatsApp, and Messenger, and is also available directly at meta.ai — no download or account is required to start a chat, though logging in with a Meta account lets you save conversations and generate images.

For developers, Meta opened a paid API for Muse Spark 1.1 in July 2026: $1.25 per million input tokens and $4.25 per million output tokens, with cached input discounted to $0.15 per million, as of July 2026 per Meta’s own developer documentation. The preview is currently limited to US-based developer accounts. Llama, by contrast, remains free to download and run yourself — you only pay for whatever compute you use to host it.

In the News

Meta’s two-track strategy played out within days this month: it opened Muse Spark 1.1’s coding-focused API to developers, and separately launched Muse Image, its first standalone image-generation model for consumers — a release that also drew criticism over how it handles people’s photos.

FAQ

Is Llama the same as Meta AI?
No. “Meta AI” is the assistant product; Llama is one of the model families that can power it. Since April 2026, the Meta AI assistant has run primarily on Muse Spark, not Llama.

Can I download Muse Spark like Llama?
No. Llama’s weights are downloadable under Meta’s license; Muse Spark is closed and only accessible through Meta’s app or its paid API.

Is Meta AI free to use?
Yes, for consumers — chatting and image generation inside the Meta AI app or Meta’s other apps costs nothing. The Muse Spark developer API is paid, billed per token.

Why did Meta build a second model family instead of just improving Llama?
Meta has said its pace of model releases lagged competitors; Muse, built by the newly formed Superintelligence Labs unit, was designed to close that gap with a model Meta fully controls end to end.