Meta's Muse Spark and the Muse Family: The Big Shift from Open to Closed

Meta's first model from its Superintelligence Labs is a multimodal reasoning model you can already use in the Meta AI app, but, unlike Llama, you cannot download it. What it does, how to reach it, and what the name means.

Sahi Padhai · 2026-09-30 · 3 min read

For years Meta was the open-weights company. Its Llama models could be downloaded, fine-tuned and run on your own hardware, which made Meta the darling of the open-source AI community. Muse Spark marks a turn.

Which Muse?

"Muse" is also used elsewhere in tech. In this article it means Meta's Muse family, starting with Muse Spark. Checked on 30 September 2026 against Meta's announcement and news coverage.

What it is

Muse Spark is the first model from Meta Superintelligence Labs, the group Meta formed to pursue what it calls "personal superintelligence". Meta describes it as a natively multimodal reasoning model with support for tool use, visual chain of thought and multi-agent orchestration. It was introduced in April 2026 and followed by a Muse Spark 1.1 update in July 2026, according to news coverage.

Unpacking those phrases:

  • Natively multimodal: it was trained on text and images together, not a text model with a vision add-on bolted on.
  • Reasoning model: it can spend extra computation "thinking" before it answers, like the long-reasoning modes of other frontier models.
  • Visual chain of thought: it reasons step by step about what it sees (a diagram, a photo of a broken appliance) and can annotate images while explaining.
  • Tool use: it can call external tools, which is the foundation of agents.
  • Multi-agent orchestration: several copies can work on parts of a problem in parallel.

Meta highlights multimodal tasks such as visual science questions and entity recognition, and health-related reasoning that it says was developed with input from more than a thousand physicians. Reports also describe a context window of up to a million tokens.

The names: Muse and Spark

Muse is the family name; Spark is the first member. Meta has said Muse Spark is "the first in the Muse family", which tells you the pattern it intends to follow: a family name, then names for individual models, probably spanning different sizes. Beyond that, the company has not published a public naming rulebook, so anything more would be guesswork.

Modes

Meta describes a standard mode plus a Contemplating mode, where multiple agents reason in parallel before answering. This is Meta's answer to "deep thinking" modes from other labs, and it trades speed for accuracy on hard problems.

How you can use it

  • Consumers: Muse Spark powers the assistant in the Meta AI app and on meta.ai, and Meta is rolling it across WhatsApp, Instagram, Facebook, Messenger and Threads, with plans for its smart glasses.
  • Developers: at launch Meta described a private API preview for selected users, with no public pricing announced. That is why this article has no price table.

The bigger story: open versus closed

The most important thing about Muse Spark is not a benchmark. It is that, unlike Llama, it is proprietary: there are no downloadable weights. Coverage describes the release as a move from an open-source approach to a product-focused, closed one.

Why should an AI learner care?

  • Open weights let you run, study and fine-tune a model yourself, which is how most people learn how these systems behave.
  • Closed models are reached only through the company's apps or API, so you use what they expose and pay what they charge.

Both approaches will keep existing. Other labs still release open models, and the gap between open and closed models is a moving target.

Muse Spark in the map

In the layers from our AI tool map, Muse Spark is a model, and the Meta AI app is the product wrapped around it. Meta has not (yet) launched a coding-agent harness like Claude Code or Codex, so there is no "Muse CLI" to compare in the terminal agents article.

What you can do with it today

  • Ask Meta AI questions that mix images and text.
  • Use it inside the apps you already have, such as WhatsApp.
  • If you are a developer, apply for the API preview and watch for public pricing.