Muse Spark Explained: Meta's AI Model in Plain English
Published 8 min read
Muse Spark is the AI model Meta built to power its chat assistant, Meta AI, and its agent app, Muse. A model is the “brain” that reads your request and writes the answer. Apps sit on top of it.
Meta launched Muse Spark on April 8, 2026. Since then, it has shipped three updates and an open version anyone can download.
Muse Spark is one model in a bigger family. For the full picture, see our guide to Meta’s Muse AI models.
This page explains where Muse Spark runs, how it scores and what it costs, in plain English.
Muse Spark key facts (as of October 26, 2026):
- Made by: Meta Superintelligence Labs, Meta’s AI research group.
- Where you can use it: meta.ai, the Meta AI app and the Muse app.
- Latest version: Muse Spark 1.3, released September 2, 2026.
- Open version: Muse Spark Open, released October 15, 2026.
- API price: $1.25 per million input tokens and $4.25 per million output tokens.
What is Muse Spark?
Muse Spark is the first model in Meta’s Muse family of AI models. Meta calls it a “natively multimodal reasoning model” in its launch post.
In plain words, it can read text, images and more. It can also think through hard problems step by step.
Meta’s announcement says it can also use tools and coordinate several AI helpers at once. Meta says it needed “over an order of magnitude less compute” than Llama 4 Maverick to reach the same level. That means more than 10 times less computing power.
Muse Spark is not the Muse app
The names can confuse people. Here’s the simple split:
- Muse Spark is the model.
- Muse is Meta’s personal AI agent app, which runs on that model.
- Meta AI is Meta’s chat assistant, which also runs on Muse Spark.
New to the app? Our guide to what Meta Muse is covers how the agent works.
Muse Spark in Meta AI and the Muse app
Muse Spark powers both Meta AI and the Muse app, so you’ve likely used it already. It came first to Meta AI. At launch, Meta said it was “available now on meta.ai and the Meta AI app.”
Meta’s newsroom post says it started in the US. Meta also planned to bring it to WhatsApp, Instagram, Facebook, Messenger, Threads and its AI glasses.
Modes in Meta AI
Meta AI offers Muse Spark in a few modes:
- Instant: quick answers.
- Thinking: more time for harder problems.
- Contemplating: several AI helpers reason side by side. Meta said it was “rolling out gradually.”
We couldn’t confirm that Contemplating mode is open to everyone today.
In the Muse app
Meta’s Muse launch post says the agent app is “powered by Muse Spark” too.
Meta hasn’t said which exact Spark version runs in Meta AI or the Muse app today.
Meta also says its bigger Muse Wonder model will “soon” power Muse, starting with subscribers. To see how the two apps differ, read our Muse vs Meta AI comparison.
Muse Spark benchmarks
Muse Spark scores well, but it isn’t the top model, and the numbers depend on who ran the test. A benchmark is a standard test that lets people compare AI models.
What Meta reports
Meta’s own numbers for Contemplating mode at launch:
- 58% on Humanity’s Last Exam, a very hard quiz across many subjects.
- 38% on FrontierScience Research, a science research test.
For Spark 1.3, Meta’s developer model page compares it with rival models. A few rows:
| Test (Meta’s numbers) | Spark 1.3 | Spark 1.2 | GPT 5.6 Sol | Opus 5 |
|---|---|---|---|---|
| OSWorld 2.0 (using a computer) | 66.9 | 47.6 | 62.7 | 68.3 |
| Terminal-Bench 2.1 (coding) | 88.8 | 82.9 | 88.8 | 86.7 |
| DeepSearchQA (research) | 90.3 | 85.9 | 93.1 | 90.4 |
These are Meta’s own numbers. Companies pick the tests they share, so treat them with care.
What an independent index says
Artificial Analysis is an independent group that scores AI models on one combined index. The catch is that it changes the index over time.
It adds tests and swaps old ones. So a score from April and a score from October aren’t comparable.
| Date | Index reading | Muse Spark score |
|---|---|---|
| April 8, 2026 | Launch article | Spark 52, 4th place at the time |
| August 5, 2026 | Newer index | Spark 1.0: 43, 1.1: 51, 1.2: 54 |
| October 26, 2026 | Index v4.3.2 | Spark 1.3: 48, Spark Open: 39 |
Notice that Spark 1.0 dropped from 52 to 43. The model didn’t get worse. The test got harder.
On the current Spark 1.3 page, it ranks 25th of 215 models. The page marks it “deprecated,” meaning outdated, and points to Muse Wonder. Spark Open ranks 11th of 121 open-weight models, well above their typical score of 19.
Muse Spark versions: 1.1, 1.2 and 1.3
Meta has released four versions of Muse Spark, plus an open one. Each one added new skills or new places to use it.
| Version | Date | What changed | Where |
|---|---|---|---|
| 1.0 | April 8, 2026 | First Muse model | meta.ai, Meta AI app |
| 1.1 | July 8, 2026 | Public API preview for US developers | Meta Model API |
| 1.2 | August 5, 2026 | Tuned for coding | Muse Code, Meta Model API |
| 1.3 | September 2, 2026 | Better at long agent tasks | Muse Code, Meta Model API |
| Spark Open | October 15, 2026 | Open weights anyone can download | Hugging Face, cloud hosts |
Muse Spark 1.1
Version 1.1 opened Spark to outside developers. Meta’s developer blog says it added web search, computer use and a setting for how hard it thinks.
Muse Spark 1.2
Version 1.2 was built for coding. It launched with Muse Code, Meta’s coding agent for developers. It can take in about a million tokens, or small chunks of text, at once.
Muse Spark 1.3
Version 1.3 is better at agent work. It can run longer tasks, ask questions when unsure and confirm before big actions.
Meta says it uses about 20% fewer tool calls and 25% fewer tokens than 1.2. It also resists prompt injection better, which is when a web page hides instructions to trick an AI.
Is Muse Spark open source?
The main Muse Spark models are not open, but Muse Spark Open is. The Spark versions Meta runs itself, 1.1 to 1.3, stay private. You can only use them through Meta’s apps and API.
At launch, Meta said it hoped to open-source future versions. It did that on October 15, 2026. Meta’s Spark Open post calls it “the open-weight version of Muse Spark.”
- Size: 816 billion parameters, with 42 billion active at a time. Parameters are the numbers a model learns.
- Context: about 1 million tokens, so it can read very long documents.
- License: the “Muse Model License,” with limited commercial terms.
- Where: Hugging Face and big cloud platforms like Google Cloud and Microsoft Foundry.
Open-weight isn’t quite the same as open source. Meta shares the model files, but the license limits some business uses.
Meta says Spark Open can run on two M5 Ultra Mac Studios. The Muse app itself still runs in Meta’s cloud.
Muse Spark API and pricing
Developers pay $1.25 per million input tokens and $4.25 per million output tokens for Muse Spark, as of October 2026. An API is how developers plug a model into their own apps. Prices come from Meta’s pricing page.
| Plan | Input (per 1M tokens) | Cached input | Output (per 1M tokens) | Trains on your data? |
|---|---|---|---|---|
| Standard (Spark 1.1 to 1.3) | $1.25 | $0.15 | $4.25 | No |
| Contributor (Spark 1.2 and 1.3) | $0.10 | $0.002 | $0.20 | Yes |
| Muse Wonder, for comparison | $4.00 | Not listed here | $20.00 | Not listed here |
A few notes:
- Contributor mode is much cheaper, but Meta uses your prompts, code, files and answers to train future models.
- Web search costs $2.50 per 1,000 searches.
- Muse Spark Open has no price on Meta’s page. You can run it yourself or use a hosting company.
- Free credits: some outlets reported $20 in credits for new accounts. We couldn’t confirm this on Meta’s page.
Where to go next
- Compare the whole family in our guide to Meta’s Muse AI models.
- Choosing an app, not a model? Read Muse vs Meta AI.
- Want to know how the agent stacks up against OpenAI? See Meta Muse vs ChatGPT.
Sources
- Meta: Introducing Muse Spark (April 8, 2026)
- Meta: Build with Muse Spark (July 8, 2026)
- Meta: Meet Muse Spark 1.2 and Muse Code (August 5, 2026)
- Meta: Introducing Muse Spark 1.3 (September 2, 2026)
- Meta: Muse Spark Open (October 15, 2026)
- Meta: Meta Model API pricing (read October 26, 2026)
- Artificial Analysis: Muse Spark 1.2 article (August 5, 2026) and Muse Spark Open page (read October 26, 2026)
Frequently asked questions
How do I use Muse Spark?
Go to meta.ai or open the Meta AI app, where Muse Spark powers Meta AI. You can also use the Muse app, Meta's personal AI agent, which Meta says runs on Muse Spark. Developers can call it through the Meta Model API, or download the open-weight Muse Spark Open model.
Is Muse Spark free?
For everyday people, mostly yes. Meta says Meta AI is free for everyday use, and the Muse app is free with a usage limit. Developers pay through the Meta Model API: $1.25 per million input tokens and $4.25 per million output tokens in Standard mode, as of October 2026.
Is Muse Spark good?
It depends on the job. Meta's own scores show Spark 1.3 close to top rival models on agent and coding tests. The independent Artificial Analysis index gave Spark 1.3 a 48 on October 26, 2026, ranking it 25th of 215 models. It's fast and cheap, but Meta's newer Muse Wonder scores higher.
Is Muse Spark related to Gemini Spark?
No. Muse Spark is an AI model made by Meta. Gemini Spark is a personal AI agent made by Google, which Google now calls Task in the Gemini app. They come from different companies and share only a similar name. One is a model, and the other is an agent you use.