Meta, from open Llama models to the new Muse family

A social media giant turned model maker

Meta Platforms is the company founded in 2004 as Facebook and renamed Meta in October 2021, headquartered in Menlo Park, California. It is not an AI lab in the strict sense, but it has run major AI research for years and designs its own large language models.

Llama: the open-weight bet

The Llama family was Meta’s AI showcase for a long time. Its models were released with open weights under a community license: anyone could download them, fine-tune them and host them, subject to certain conditions. Llama became the starting point for countless projects, local assistants and derivative models.

2026: the Muse turn

In April 2026, Meta Superintelligence Labs, a team set up to rebuild the group’s AI efforts from the ground up, introduced Muse Spark, the first model in a new family. It is a multimodal reasoning model that can use tools and coordinate several agents. Unlike Llama, it is proprietary: its weights have not been released.

Meta then expanded the Muse lineup (Wonder, Glimmer, plus image and transcription models) and released Muse Spark Open, an open-weight version available on Hugging Face. The result is now a mixed policy: closed models at the top, open models alongside them.

How to access Meta’s models

• Meta AI, the consumer assistant: see the Meta AI page.
• The Meta Model API, which gives access to the Muse models.
• Downloading the weights of Llama and Muse Spark Open, to host them on your own infrastructure.

What sets Meta apart

In practice, Meta’s strength is distribution: its models reach people directly inside apps they already use. The downside is a change of course that means checking, model by model, what is open and under which license. To compare it with other open-model makers such as Mistral, see the open source models page.