# What is Muse Spark, Meta's first model from its superintelligence lab?

Date: 2026-04-09 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-meta-muse-spark/

**Answer:** Muse Spark is a new AI model that Meta announced on April 8, 2026, the first from its Meta Superintelligence Labs team and the start of a new "Muse" series that will replace Llama inside Meta AI. It is small and fast by design, understands images as well as text, now powers the Meta AI app and website in the US, and unlike Llama it is not being released openly for now.

Yesterday, Wednesday, April 8, 2026, Meta announced **Muse Spark**, the first AI model from **Meta Superintelligence Labs**, the team it built last year to get back into the front rank of AI. Muse Spark already runs the Meta AI app and the meta.ai website, and Meta says it will reach WhatsApp, Instagram, Facebook, Messenger and its AI glasses in the coming weeks.

It is the first new model Meta has released in about a year, The Guardian notes, and the first since its Llama 4 models disappointed many observers. Muse Spark begins a new family: Meta calls it the first of the **Muse** series and plans larger versions.

## What Meta released

Meta says its new lab spent nine months rebuilding its AI technology from scratch. Muse Spark is the first result, and Meta stresses that it is deliberately small and quick, a starting point rather than its largest model. The Guardian reports the series goes by the internal name Avocado.

Alongside the model, the Meta AI assistant got a new look and new features:

| Feature | What it does |
|---|---|
| Instant and Thinking modes | Quick replies, or slower step-by-step reasoning for harder questions |
| Parallel subagents | Several helper agents work on parts of one question at the same time |
| Seeing images | Understands photos, such as a shelf of snacks or a product you scan |
| Health answers | More detailed replies on common health questions, built with input from physicians |
| Visual coding | Builds simple websites and mini-games from a description |
| Shopping mode | Suggests clothes, decor or gifts, drawing on creators and brands across Meta's apps |

The new features start in the United States. Developers get only a **private preview** through an API for selected partners, and Meta says it hopes to open-source later versions.

## How it works

**Multimodal** means a model handles more than one kind of input. Muse Spark reads text and also looks at images, so you can show it what you mean instead of describing it. Meta's example: photograph the snacks at an airport shop and ask which have the most protein.

**Subagents** are the other key idea. An agent is an AI that carries out a task in steps; a subagent is a helper it sends off to handle one part. Meta's example is planning a family trip to Florida: one helper drafts the itinerary, another compares Orlando with the Keys, and a third looks for activities for children, all at once. The Guardian describes an extra Contemplating mode that runs several agents together to reason harder, aimed at the same kind of deep work as rival "thinking" products.

The biggest change for Meta is how the model is shared. Llama models were **open-weight**: Meta published the trained numbers that make up the model, so anyone could download and run them. Muse Spark is **closed**: you use it only through Meta's services, and Meta has not disclosed its size.

## How it compares

Meta calls Muse Spark its most powerful model yet. Independent testers quoted by The Guardian found a mixed picture: close to the leading models from Google, OpenAI and Anthropic in language and understanding images, but behind them in coding and abstract reasoning. It tied for fourth place on a combined index from the evaluation firm Artificial Analysis.

Alexandr Wang, who leads the lab, acknowledged on social media that the model still has rough edges to polish and said bigger versions are coming, some of them to be released openly. Mark Zuckerberg had told investors in January to expect good early models whose main value would be showing how fast the team is improving.

## A worked example: using photos and subagents together

A request that plays to Muse Spark's strengths combines a picture with a task that splits naturally into parts:

```prompt
[Photo of my living room]
I want to redecorate this room for under $500. Suggest three style
directions that suit the light and furniture shown. For each, list
five items to buy and explain where each one would go in the photo.
```

1. The model studies the photo: window position, sofa color, floor.
2. Helper agents can explore each style direction in parallel.
3. Shopping mode can suggest items, which you should still compare on price and reviews yourself.

> **Try it:** Take a photo of the inside of your fridge and ask Meta AI for two dinner ideas using only what it can see. Check whether it identified the items correctly before you trust the recipe.

## Why it matters

Meta has spent heavily on this team, including a $14.3 billion deal that brought in Wang from Scale AI, and Muse Spark is the first public test of that spending. The early verdict is "competitive, not leading," which matches what Zuckerberg told investors to expect.

It also marks a turn in strategy. Meta was the best-known champion of open AI models through Llama. Its newest model is closed, at least for now, while Meta says it still hopes to release future versions openly. And because Meta's apps reach more than 3.5 billion people, a better assistant there could quickly become the AI that many people use most, often without seeking it out.

## How can you try Muse Spark, and what should you watch for?

If you are in the United States, open the Meta AI app or go to meta.ai; Muse Spark is the model behind them now. Try both Instant and Thinking modes, and test the photo features with everyday scenes. Outside the US, Meta says the new modes will spread to more countries in the coming weeks.

Watch for three things. First, accuracy on health questions: treat answers as general information, not a diagnosis. Second, shopping suggestions: they come from inside Meta's own ecosystem, so compare prices elsewhere. Third, privacy: think about what your photos show before you share them.

Builders who relied on downloading Llama should note that Muse Spark is not available that way. If you need an open model you can run yourself, Llama remains an option while Meta decides how much of the Muse series it will open.

## Sources

- Introducing Muse Spark: MSL's First Model, Purpose-Built to Prioritize People: https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/
- Meta debuts new AI model in first test of costly 'superintelligence' team: https://www.theguardian.com/technology/2026/apr/09/meta-first-ai-model-muse-sparks