Muse Spark Open Explained: Meta's New Open-Weight Model for Coding and AI Agents
Quick answer: Muse Spark Open is Meta's new open-weight AI model, announced October 15, 2026 and described by Meta as the most capable U.S. open-weight model to date. It has 816 billion total parameters with 42 billion active at a time, a 1-million-token context window, and native multimodal design tuned for complex coding and agentic work. It's free to download, with a license that allows limited commercial use — but it's a developer model, not the Muse assistant app most people use.
Meta is having a busy October. On October 12 the company announced Muse Wonder and Muse Wonder Full, frontier models strengthening the Muse personal agent. Three days later, as reported this week by TechCrunch, MacRumors and Reuters, Meta unveiled something aimed at a different audience entirely: Muse Spark Open, an open-weight model for developers.
If the names are starting to blur together, here's the simple map: Muse is the personal assistant on your phone (launched September 8, 2026). Muse Wonder and Muse Wonder Full are the frontier models making that assistant smarter. Muse Spark Open is the developer-facing heavyweight — a model you can actually download and run yourself. Well, yourself plus about a terabyte of RAM. More on that in a minute.
| Detail | Muse Spark Open |
|---|---|
| Total parameters | 816 billion |
| Active parameters at a time | 42 billion |
| Context window | 1 million tokens |
| Design focus | Complex coding tasks and agentic work |
| Multimodality | Natively multimodal |
| Cost | Free to download; license allows limited commercial use |
| Where to get it | APIs: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI. Weights: Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, Microsoft Foundry Models |
What "open-weight" actually means — and why it isn't "open source"
People often use the two terms interchangeably, and they're close but not the same. An open-weight model publishes its trained parameters — the neural-network weights — so developers can download, run, fine-tune and build on them. It is not necessarily open source: Meta is not handing over its training code, training data or full training pipeline the way a classic open-source project would.
That distinction matters in practice. With the weights downloadable and modifiable, a company can take Muse Spark Open, adapt it to its own code base or internal tools, and run it on its own infrastructure — without depending on Meta's API forever. The license keeps some limits: it's free to download and allows limited commercial use, so read the terms on Hugging Face or Meta's release page before building a product on it.
This isn't Meta's first open-weight release of the year either. Back in August 2026 the company released Muse Glimmer, a smaller open-weight model. Spark Open is the big sibling — and according to Meta, the most capable U.S. open-weight model to date.
What it's built for: hard coding and agentic work
Muse Spark Open is not a general-purpose chat companion. Meta describes it as natively multimodal and built for complex coding tasks and agentic work — the kind of jobs where a model doesn't just answer a question but executes a multi-step workflow.
To show what that looks like, Meta paired the model with "Muse Code," its AI coding terminal, and demonstrated agentic workflows including coding fixes for a Mastodon client and reviewing plus cropping footage inside DaVinci Resolve. In other words: the demo wasn't a chatbot conversation, it was the model doing work across real developer tools — writing code, then working with video footage. The 1-million-token context window is the enabler: it lets the model hold an entire code base or long session in memory while it works.
The Muse personal assistant is optimized to help you get things done; Muse Spark Open is optimized so developers can build things that get things done.
The Mac connection: designed with Apple, needs serious RAM
Here's the most surprising part of the announcement. Meta designed Muse Spark Open with Apple to run on Mac hardware — but "run on a Mac" here means something very specific. The model needs at least around 1TB of RAM, and Meta says it runs best on two 512GB Apple M5 Ultra Mac Studios, each with 512GB of unified memory and 1.2 TB/s of memory bandwidth, using the latest MLX framework.
Two things stand out about that setup. First, the 512GB Mac Studio configuration isn't even released yet — it's expected later this month. Second, the current 256GB Mac Studio already costs over $10,000, so the "best" local setup for this model would be two machines in that league, using a configuration you can't buy yet.
That's just the honest math of 816-billion-parameter models: they need enormous memory, and Apple Silicon's unified memory architecture is one of the few ways to get it in a desktop. But it means that for most developers, "running Muse Spark Open" will really mean hitting an API — Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter or Together AI — rather than running the weights locally. Which is fine: Meta clearly built it for that path too.
How to actually try it: APIs and downloadable weights
As of the announcement, Muse Spark Open is available now through a long list of API platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Prefer to host it yourself? The weights are downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.
The same month also brought Muse Wonder and Muse Wonder Full, the frontier models strengthening the Muse personal agent people use on their phones. Spark Open is the developer side of that strategy: while Wonder-class models make the consumer product smarter, Spark Open lets the broader ecosystem build with Meta's best weights.
Frequently asked questions
Is Muse Spark Open really free?
Yes, the model weights are free to download, and the license allows limited commercial use. Running it is the expensive part — either the hardware (around 1TB of RAM, ideally two 512GB M5 Ultra Mac Studios) or compute on an API platform, where you'll pay for usage according to each provider's pricing. Check the pricing page of whichever platform you use before scaling up.
Can I run Muse Spark Open on my laptop?
Almost certainly not. It needs at least around 1TB of RAM; Meta says it runs best on two 512GB M5 Ultra Mac Studios (512GB unified memory each, 1.2 TB/s bandwidth) using the latest MLX framework. For nearly everyone, accessing it through an API provider like Fireworks AI, Groq, Together AI or OpenRouter is the realistic route.
What's the difference between Muse Spark Open and the Muse assistant app?
They're different products in the same family. The Muse personal assistant (launched September 8, 2026, available in the US and Canada) is the app that does tasks for you. Muse Spark Open is an open-weight developer model for building AI applications, coding workflows and agents — you wouldn't "chat" with it the way you chat with the assistant.
What is Muse Glimmer?
Muse Glimmer is a smaller open-weight model Meta released in August 2026, two months before Spark Open. Think of it as the lighter sibling: Spark Open is the flagship 816-billion-parameter release, while Glimmer targets lighter-weight use cases.
If you're not a developer and just want the personal agent — something that books, buys and files things for you — that's the Muse assistant, and getting set up properly matters: a Meta account, 18+ verification, and the right plan before handing an agent access to your accounts. We offer a Muse account registration assistance service that walks you through creating your own account for personal use, per platform terms. If you'd rather have help getting your own Muse account set up correctly than puzzle through the verification steps alone, our Muse registration help page has the details.