What Is Muse Spark? Meta’s AI Model, Its Agent App, and What They Do

In the space of five months Meta has shipped a model, three updates to it, a coding tool, a small open model and an agent app, and several of them share a name. If you have seen “Muse” and “Muse Spark” in the same week and assumed they were the same thing, that is a reasonable mistake. They are not. This post sorts out what each one is, what the model is good at, where it is behind, and what any of it means if you use AI at work.

The short answer

Muse Spark is a large language model. It is the first model from Meta Superintelligence Labs, the group Meta set up to build its frontier AI, and it was launched on 8 April 2026. It is the engine. You do not usually “use Muse Spark” directly; you use a product that runs on it.

Several products run on it, and this is where the names get confusing:

• —The Meta AI app and the meta.ai website are Meta’s chatbot. Since April they have been powered by Muse Spark. This is the one most people will meet first.
• —Muse is a separate app: a personal AI agent that does tasks for you rather than only answering questions. Meta introduced it on 8 September 2026. It also runs on Muse Spark.
• —Muse Code is a coding agent that runs in a developer’s terminal, released in beta in August and out of beta since 31 August.
• —Muse Glimmer is a much smaller open-weight model that anyone can download and run themselves.

So “Muse vs Muse Spark” is not really a comparison. Muse is the app; Muse Spark is the model inside it. The practical guide to the chatbot side is how to use Muse Spark, and the agent app has its own walkthrough in setting up Meta’s Muse app.

Why it matters that this is not Llama

Meta’s earlier models, the Llama family, were released with their weights, which meant companies could download and run them on their own machines. Muse Spark is proprietary. Meta’s own announcement describes it as a new direction for Meta’s AI work, and at launch the only way for developers to reach it was a private API preview.

That may change. When 1.3 shipped, Mark Zuckerberg promised open weights of Muse Spark “soon” without giving a date. Until that happens, treat it like the other big commercial models: you use it through Meta’s products or Meta’s API, on Meta’s terms.

The version timeline, April to September

The dates below are from Meta’s own posts and the coverage around them, and are accurate as of September 2026.

1. 018 April: Muse Spark launches in the Meta AI app and on meta.ai. It handles text and images, can use tools, and offers Instant and Thinking modes, plus a “Contemplating” mode where several agents reason in parallel.
2. 029 July: Muse Spark 1.1 and a public preview of the Meta Model API, with a 1 million token context window, computer use, support for MCP servers, and reading of images, video and PDFs.
3. 035 August: Muse Spark 1.2 arrives alongside Muse Code, the terminal coding agent.
4. 0410 August: Muse Glimmer, a 30-billion-parameter open-weight model under an Apache 2.0 licence, built for running agents locally.
5. 052 September: Muse Spark 1.3.
6. 068 September: the Muse agent app, running 1.3, rolls out in the US.

The 1.3 update is the one most relevant to everyday use. Meta says it uses about 20% fewer tool calls and about 25% fewer tokens than 1.2, is better at long multi-step tasks, and resists prompt injection better. Meta also says it now asks a clarifying question when a request is ambiguous and checks with you before doing something consequential.

What it is good at

Start with the independent numbers, and hold them loosely. On the Artificial Analysis Intelligence Index, Muse Spark 1.3 scored 61 in its public setting, level with GPT-5.6 Sol at max and Grok 4.6, and a few points behind the top Claude models. Of the leading models in that analysis, it was the cheapest to run through the test, and it did well on long-context tasks.

In the chat product, the useful part is the tools. An early hands-on review found meta.ai could search the web and open pages, run Python in a sandbox for data work and charts, find and count objects in images, generate images, and search public Instagram, Threads and Facebook posts. That last one matters if your job touches social media.

Meta also leaned on health. It says its health answers were trained with input from more than 1,000 physicians, which is worth knowing, and not a reason to skip a doctor.

Where it is behind

Meta was unusually direct about this at launch. It said Muse Spark still had gaps in long-horizon agentic systems and coding workflows. Version 1.3 claims progress on long tasks. Even so, Artificial Analysis recorded small drops on knowledge tests and about four points lost on a legal reasoning test compared with 1.2.

Two more things are fair to weigh. First, the safety testing: Apollo Research, which tested the model, reported the highest rate of “evaluation awareness” it had seen, meaning the model seemed to notice when it was being tested. Apollo judged this not serious enough to block release, but it is the kind of finding worth knowing about.

Second, the company. TechCrunch’s launch coverage of Muse points out that Meta paid a $5 billion FTC penalty in 2019, and recalls Cambridge Analytica and passwords once stored in readable form. Meta says Muse keeps your conversations and connected data away from its ad systems, and you can opt out of training. Whether that is enough is your call, and whether Meta’s Muse is safe to connect to your accounts goes through the permissions one by one.

None of the benchmark numbers above are your work. A score measures a test set. It does not tell you whether the model will get your client’s figures right, which is why the habits in what AI is actually bad at apply to Muse Spark exactly as they apply to everything else.

A worked example

Say you run the events budget for a small team and you have a spreadsheet export of last quarter’s spending. Here is how the same job looks in each product.

In the Meta AI chat, you upload the file and ask: “Group these costs by supplier, show the three biggest, and draw a bar chart. Tell me which rows you could not categorise.” The last sentence is doing real work. It turns a confident summary into one that admits its gaps. If you want the prompt tighter, writing a prompt that works on the first try covers the pattern.

Then you check it. Pick the biggest supplier and add up its rows yourself. If your total matches, trust the chart more. If it does not, you have learned something useful about the tool before it cost you anything. The routine for checking without being the expert is in how to check an AI answer when you are not the expert.

In the Muse app, the job is different. You would ask it to act: email the top three suppliers for next quarter’s quotes and add the replies to a list. That means connecting your email. Meta says you choose how much access each app gets, read-only or able to send, and that Muse asks before sending an email or making a purchase. Start read-only. Give it the send permission only once you have watched it draft a few emails you would have sent yourself.

The model is the same in both. What changes is whether it hands you an answer to check, or takes an action you have to approve.

What it means if you use AI at work

Muse Spark is now a serious model, close to the leaders on independent tests and cheap to run through the API. For most people that does not mean switching. It means there is one more capable assistant, and one more agent app asking for access to your inbox. Comparisons are in Muse Spark vs ChatGPT if you are deciding between them. If you live in Google’s apps, Muse Spark vs Gemini is the closer match.

The bigger shift is the move from chat to agents. Muse, Google’s Gemini Spark and Anthropic’s Claude Cowork all do tasks, not just answers. The skill that transfers between them is knowing which of your tasks are safe to hand over, and finding the repetitive part of your work is the place to start.

What to do Monday

1. 01Open meta.ai or the Meta AI app and run one real, low-stakes task you already know the answer to. Note where it is right and where it guesses.
2. 02Try the same prompt in the assistant you already use. The difference tells you more than any benchmark.
3. 03If you try Muse, connect one service, read-only, and read every approval request before you tap it.
4. 04Check your settings for the option to keep your conversations out of Meta’s model training, and decide deliberately.
5. 05Write down one task each week that an agent could take off your plate, and one it should never touch.

Model names will keep changing; the judgement about what to hand over will not. Coursium is an iPhone app that teaches people to use AI at work, including how to check an output and when not to trust one. Stay ahead of AI and learn the tools on your phone.