← Labs & modelsLab
Meta.
Menlo Park, USMeta Superintelligence Labs formed 2025Models: Some online only, some downloadable
In plain English
The company behind Facebook, Instagram and WhatsApp, which has been building its own AI for years. Its earlier models were called Llama, and its newest family is Muse. You can reach Muse through all of those apps, or directly, including through Muse Charm, a small keyring device due out for the 2026 holidays. Some of its models, like Llama, you can download and run yourself.
Models to know
The Muse family runs from frontier models like Muse Wonder through to Muse, a personal AI agent. Meta has even put Muse in a cute Tamagotchi-style keyring called Muse Charm. If you want to run a model yourself, look at Llama 4 Scout or Muse Spark Open.
Who's in charge
Mark Zuckerberg, founder and CEO. Alexandr Wang leads its AI lab, Meta Superintelligence Labs.
Things to be aware of
In April 2026 Meta started recording US employees' keystrokes, mouse clicks and screenshots to train its AI agents. It paused the programme in June 2026. In 2025, Reuters also found Meta's internal rules had allowed its AI chatbots to have romantic conversations with children. Meta removed those rules and added safeguards for teenagers.
Read more
Fortune on Meta tracking employee keystrokes · Reuters on Meta's chatbot rules
Models
In plain English
Meta's strongest model for long-running personal and work agents.
Good for
Complex software work, assistants that remember a lot, and tasks that carry on across many steps.
Things to be aware of
The more memory, tools and time an agent has, the more it can do and the more it can get wrong. Tighten permissions and checks as it gets more capable.
- Context window
- 1.0M tokens
In plain English
Meta's main model for coding and agent work at scale.
Good for
Coding, reasoning over long documents, and repeated tool use.
Things to be aware of
Meta offers a cheaper version that may use your content to improve its products, so check which one you're buying.
- Context window
- 1.0M tokens
In plain English
A downloadable version of Muse Spark, with 816 billion parameters.
Good for
Organisations that want more control over where it runs and where their data goes.
Things to be aware of
Open weight doesn't mean open source, and even the smaller version is about 537GB. Running it is a serious infrastructure job.
- Context window
- 1.0M tokens
In plain English
Meta's downloadable model that works with text and images, with a very large context window.
Good for
Running it yourself, and work that needs very long documents.
Things to be aware of
A 10-million-token context window doesn't mean it uses all of it well. And its licence isn't a standard open-source one.
- Context window
- 10.0M tokens
- Knows up to
- August 2024
