Muse, Spark, Wonder, Glimmer: Meta's AI family sorted out simply

Why everything is called Muse

Until 2025, Meta's AI models were called Llama and its chat assistant Meta AI. After restructuring its AI division into Meta Superintelligence Labs, Meta introduced a new brand: Muse. It now stands both for products people use directly and for the technology behind them.

It is as if a carmaker gave a car and several engines the same name. To place news correctly, a simple distinction helps:

Products are what you open and operate. At Meta, these are mainly the agent app Muse and the programming tool Muse Code.

Models are the actual AI programs that understand, plan and answer. They sit inside the products or are built into other applications by developers through an interface. At Meta, they carry an extra name: Spark, Wonder, Wonder Full and Glimmer.

So when a news item says Muse has improved, it is worth checking whether it means the app or a new model.

The products: the Muse app and Muse Code

The Muse app is Meta's personal AI agent. Meta launched it on 8 September 2026. You write to it as you would to a person, in its own app or in WhatsApp, and it acts by itself: answering emails, booking tables, shopping, filling in forms. According to Meta, each user gets their own isolated computer in Meta's cloud for this. The basic version is free, and there are subscriptions for more usage. According to CNBC, they cost 20 or 100 US dollars a month. The app has so far run on the Muse Spark model. Meta does not offer it in Europe.

Muse Code is Meta's programming tool, comparable to Anthropic's Claude Code or OpenAI's Codex. It is aimed at developers and at people who want to build software with AI help. Since 12 October, Meta's strongest model, Muse Wonder Full, is also available here, in the subscriptions for 50 or 200 US dollars a month. According to VentureBeat, data entered in these subscriptions may be used for training by default; you can switch this off for future submissions.

There are further Muse products, such as a version for small businesses that answers customer enquiries and creates content, and the image model Muse Image, which generates and edits images.

The models: Spark, Wonder, Wonder Full and Glimmer

Muse Spark is the first model from the new division. Meta introduced it in April 2026 and has since improved it up to version 1.3. It is comparatively cheap: developers pay 1.25 US dollars per million input tokens and 4.25 US dollars per million output tokens. Spark has so far powered the Muse app.

Muse Wonder has been Meta's new, larger model class since 12 October 2026. It is meant to remember important things across several conversations and stick with difficult tasks. Meta calls this Active Memory and Active Focus. Through Meta's interface, Wonder costs 4 and 20 US dollars per million tokens, and it is due to come to the Muse app soon, starting with paying users.

Muse Wonder Full is the stronger variant of Wonder and Meta's most capable model so far. It is currently only available in the Muse Code subscriptions, not through the interface.

Muse Glimmer is the small-format counterpart. Meta released it in August 2026, with 30 billion parameters and under the permissive Apache 2.0 licence. Anyone may download and use it, including commercially. Meta trained it on the answers of Muse Spark, a technique called distillation. Glimmer is designed to run as an agent on your own Mac or PC, even without internet. Compressed, it needs under 20 gigabytes of memory according to Meta, so it fits on computers with 24 or 32 gigabytes of graphics or main memory. Meta does not make the Wonder models available for download.

How strong are the models in comparison?

The independent testing firm Artificial Analysis compares AI models with an Intelligence Index made up of ten tests covering reasoning, knowledge, coding and office work. It shows how quickly Meta has caught up. Muse Spark 1.3 scored 48.1 points. Six weeks later, Muse Wonder reaches 52.4 and Muse Wonder Full 55.8 points. For comparison: the current best model, Anthropic's Claude Opus 5.5, scores 57.6, OpenAI's GPT-6 Astra 52.7.

More performance costs more, though. Artificial Analysis calculates what an average test task costs, because models use different amounts of text. With Muse Spark 1.3 it was 1.60 US dollars, with Muse Wonder 4.52 US dollars, almost three times as much. There is no such figure for Wonder Full, because Meta does not sell it per token.

For Muse Glimmer, Meta only gives comparisons with open models of the same size, such as Google's Gemma 4 and Alibaba's Qwen 3.6, against which it does well on several agent tests. It cannot be compared with the large cloud models. In return, it costs nothing on your own computer except electricity.

What can you use in Germany and the EU?

The Muse app: no. Meta only offers it in the US and has given no date for Europe. Observers cite the General Data Protection Regulation, the AI Act and the Digital Markets Act, which places special obligations on Meta as a so-called gatekeeper, as possible reasons. Meta's existing chat assistant Meta AI in WhatsApp, Instagram and Facebook is available here, however.

Muse Code and the interface for Spark and Wonder are aimed at developers. Whether your country is supported becomes clear when you sign up. Check the data-use terms before taking out a subscription.

Muse Glimmer: yes. The weights, meaning the actual model, are freely available on the Hugging Face platform. At release, Meta announced that programs such as Ollama and LM Studio would support Glimmer. It then runs entirely on your computer, and your data does not leave the device. You need a powerful computer for this, such as a Mac with plenty of memory or a PC with a large graphics card, and some interest in technology. How local AI models work in general is explained in a separate article.

In short: if you want to try Meta's new agent generation in Germany or the EU, Glimmer on your own computer or Muse Code are the most likely routes for now. The everyday app Muse with Wonder comes to the US first.