Muse Spark is Meta’s closed, multimodal AI model. It processes text, images and speech, and it has powered the Meta AI assistant since 8 April 2026. The model is the first from Meta Superintelligence Labs and succeeds Llama as Meta’s main model.
Among AI models, Muse Spark is the newcomer of 2026. Since April it has been released in four versions, of which 1.3 is the latest, and it also runs Muse, Meta’s personal AI agent.
Muse Spark is not open source and is available through meta.ai, the Meta AI app and the Meta Model API. Meta has not announced a launch in the Netherlands or the rest of Europe, which also affects whether you can use the model for business.
Who made Muse Spark?
Muse Spark was made by Meta Superintelligence Labs, the AI division Meta set up in mid-2025 under the leadership of Alexandr Wang. According to Meta, that team rebuilt the entire technical foundation for training models in nine months.
Meta deliberately describes Muse Spark as small and fast. According to Meta’s announcement, it matches the capabilities of Llama 4 Maverick using less than a tenth of the computing power. On 12 October 2026 Meta released Muse Wonder and Muse Wonder Full, which Meta calls its most capable models so far.
Mark Zuckerberg calls the goal behind Muse Spark ‘personal superintelligence’: an AI assistant geared to the personal goals and preferences of ordinary users, rather than primarily to businesses.
What can Muse Spark do?
Muse Spark processes text, images and speech, works through problems in several steps and carries out tasks using tools and subagents.
Muse Spark works in three thinking modes, which determine how much computation the model puts into an answer:
- Instant gives quick answers to simple questions, much like an ordinary chat assistant.
- Thinking works through a question step by step before answering, for questions with several parts.
- Contemplating puts several subagents to work at the same time for complex scientific or planning questions.
With image recognition, Muse Spark answers questions about a photo, compares products and helps with visual programming, such as building a website or a simple game. For health questions, Meta says it developed training data with more than 1,000 doctors.
Since July 2026, Meta AI with Muse Spark has also carried out tasks: making plans, connecting email and calendar, and putting together presentations. At launch, Meta itself acknowledged that the model lagged behind in programming and in long agentic tasks, where a model carries out many steps in a row on its own.
Which versions of Muse Spark are there?
Four versions of Muse Spark have been released, the latest being Muse Spark 1.3 of 2 September 2026. The table below lists the release date and the main change for each version, according to Meta’s announcements.
| Version | Release date | Main change |
|---|---|---|
| Muse Spark | 8 April 2026 | First model from Meta Superintelligence Labs, powering Meta AI |
| Muse Spark 1.1 | 9 July 2026 | Context window of 1 million tokens, better at tool use and computer use, public preview of the Meta Model API |
| Muse Spark 1.2 | 5 August 2026 | More training on programming tasks, launched together with Muse Code |
| Muse Spark 1.3 | 2 September 2026 | Better at agentic work, about 20% fewer tool calls and 25% fewer tokens when programming |
Muse Code, launched with Muse Spark 1.2, is a programming agent that runs in the terminal and carries out tasks in large codebases. That makes Muse Code Meta’s answer to Claude Code, Anthropic’s terminal agent that does the same kind of programming work.
According to Meta, Muse Spark 1.3 is also more resistant to prompt injection, an attack in which hidden instructions in a text or web page try to take over the model.
What is the difference between Muse and Muse Spark?
Muse Spark is the AI model. Muse is the personal AI agent that Meta launched on 8 September 2026 and that runs on Muse Spark. Muse carries out instructions, such as sending emails or booking trips, and asks for access to things like your email and calendar to do so.
At its launch on 8 September 2026, the Muse agent was only available in the United States, via iOS, Android and muse.ai. Since 27 October 2026, Muse has also been available in Singapore. Meta offers a free version and two paid subscriptions. According to Meta, Muse does not see passwords or payment details, and you can switch off the use of your conversations for training.
Muse is also the family name of all of Meta’s new models. Besides Muse Spark there is Muse Glimmer, an open model for use on your own hardware, and Muse Code for programming work, with Muse Image and Muse Video for images and video. Since 12 October 2026 the family has also included Muse Wonder and Muse Wonder Full. Muse Wonder is available as a preview through the Meta Model API and in Muse Code, and Muse Wonder Full is only for Muse Code subscribers.
Is Muse Spark open source?
No, Muse Spark is not an open-source model: Meta does not release the model weights, so you cannot download and run it yourself. At launch, Meta wrote that it wants to make future versions open. On 15 October 2026 Meta released Muse Spark Open, a separate open-weight model with 816 billion parameters. Muse Spark itself remains closed.
With Muse Spark, Meta breaks with the approach of Llama, the open-weight model family that Meta has released since 2023 and that can still be downloaded. Llama 4 remains available, but Meta has not published a new Llama generation since April 2025.
If you are looking for an open model from Meta, Muse Glimmer and Muse Spark Open are the alternatives. Muse Glimmer appeared on 10 August 2026 under the Apache 2.0 licence and, according to Meta, runs on a single consumer GPU. Muse Spark Open is available with open weights on Hugging Face, under Meta’s Muse Model License. Muse Spark itself is therefore clearly on the closed side of the divide between open-source and closed AI models.
How do you use Muse Spark?
To use Muse Spark, you choose between the consumer route through Meta AI and the developer route through the Meta Model API:
- Consumers in the United States use Muse Spark free of charge with a Meta account via meta.ai, the Meta AI app and Meta AI in WhatsApp, Instagram, Facebook and Messenger.
- Developers use Muse Spark through the Meta Model API at dev.meta.ai, and for programming work through Muse Code.
The Meta Model API charges per processed token. Meta is also rolling Muse Spark out on the Ray-Ban Meta and Oakley Meta smart glasses, starting in the United States and Canada.
Is Muse Spark available in the EU?
No, Meta has not announced Muse Spark for the Netherlands or the rest of the European Union. At the April 2026 launch, Meta announced a start in the United States, with expansion to ‘more countries’ but no list of countries or date.
Meta described the agentic features of July 2026 as available in ‘select markets’. The Muse agent can be used in the United States and, since 27 October 2026, in Singapore. Meta has not given a date for the EU. In August 2026, developers reported that the Meta Model API returns a message from Europe saying the service is not available in their region.
Meta AI in WhatsApp does work in the Netherlands, but Meta does not say which model powers that European version. So if you specifically want to test Muse Spark, you cannot be sure the European version of Meta AI will give you that.
How does Muse Spark compare with ChatGPT and Claude?
Muse Spark scores close to the models behind ChatGPT and Claude in reasoning and image recognition, but at launch it lagged behind in programming. On the Artificial Analysis Intelligence Index of April 2026, it scored 52 points, against 57 for GPT-5.4, the model behind ChatGPT at the time.
Anthropic’s Claude Opus 4.6 scored 53 points on the same index and Gemini 3.1 Pro 57. On Terminal-Bench 2.0, a test of programming tasks in the terminal, Muse Spark scored 59 points and GPT-5.4 75.1. For comparison, Llama 4 Maverick scored 18 on the Intelligence Index.
Those figures are a snapshot. OpenAI announced GPT-6 Astra on 3 September 2026 and released GPT-6 Sol and Luna on 22 September, the same day Anthropic released Claude Opus 5.5. GPT-6.1 Sol followed on 29 September 2026 in Work and Codex, and since 7 October 2026 OpenAI has been rolling out GPT-6 Sol as the model behind ChatGPT on Plus, Pro, Business and Enterprise, and GPT-6 Luna on Free and Go. Meta itself released Muse Wonder on 12 October 2026. For users in the EU, the difference therefore lies less in benchmark points and more in availability: ChatGPT and Claude are readily available here, Muse Spark is not.
If you are choosing a model for daily work, weigh benchmark points against availability, language support and privacy terms.
Is Muse Spark suitable for business use?
No, Muse Spark is not yet suitable for business use for organisations in the EU, because the model is not officially available here and Meta has not published European business terms for it.
Meta mainly aims Muse Spark at consumers: shopping, health questions, planning trips and personal tasks. The Muse agent asks for access to email and calendar, and the Meta AI assistant runs inside apps your team already uses for personal purposes.
That is exactly where the business risk lies. Employees who paste customer information into Meta AI in WhatsApp are using an AI tool without any agreements on data or oversight, a form of shadow AI that stays out of the organisation’s sight.
An internal AI policy sets out which tools your team may and may not use for client work, which data must never end up in a closed consumer model such as Muse Spark or Meta AI, and who checks the output. How to draw up such a policy is covered in the AI Literacy Course.
Frequently asked questions about Muse Spark
The remaining questions about Muse Spark are mostly about cost and about what Meta does with your data.
What does Muse Spark cost through the Meta Model API?
Through the Meta Model API, Muse Spark costs $1.25 per million input tokens and $4.25 per million output tokens at the standard rate, according to Meta’s announcement of 8 July 2026. The contributor variants of versions 1.2 and 1.3 are cheaper ($0.10 input and $0.20 output), but Meta uses that data to improve its products. You pay per use, with no fixed monthly fee. During that public preview the API was only open to developers in the United States, so you cannot simply use those rates from the EU.
Does Meta use your questions to Meta AI to train its models?
Yes, Meta uses your interactions with Meta AI, such as questions and instructions, to train and improve its AI models. For adults in the EU, Meta also uses public posts and comments on its platforms. Meta says it does not use the content of private messages with friends and family, unless someone in that conversation shares messages with Meta AI. You can object at any time through a form that Meta links to in its notifications.
