Meta just released the weights of Muse Spark Open, an 816-billion-parameter model you can download and run yourself. It is built for coding and agent work, and it is not Meta’s best model. That one, Muse Wonder, stays closed.
Here are the numbers that matter before you spend any time on it.
- Released October 15, 2026, with downloadable weights.
- 816B total parameters, 42B active per token, 1-million-token context.
- Takes text, images, video and audio in; outputs text.
- Released under the Muse Model License, which has limited commercial terms.
The short version: it is real, it is open-weight, and it is huge. Most people will use it through a hosted service, not their own hardware.
What Muse Spark Open actually is
It is a mixture-of-experts model. That means only a slice of it, about 42 billion parameters, works on each token, even though the full file is 816 billion. This keeps responses faster than a dense model of the same size.
Meta says it targets long coding tasks, multi-step reasoning and agent workflows. According to The Next Web, Meta admits it is less capable than Muse Wonder, and Alexandr Wang said the company will generally not open-source its most powerful models.
If you missed our earlier look at Meta’s first frontier model, Muse Wonder, that is the closed sibling this release sits below.
Where to get Muse Spark Open
You have two routes: grab the weights, or call it through someone else’s servers. Pricing for the hosted options was not listed in the launch coverage, so check each provider before you commit.
| Route | Where | Best for |
|---|---|---|
| Download weights | Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models | Teams that want control over data and setup |
| Hosted API | Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI | Developers who just want to try it quickly |
Why this matters: seven hosted providers on day one means you can test it without buying a single GPU.
Is the Muse Spark Open license really free?
Mostly, but read it first. OfficeChai reports it ships under the Muse Model License, which Meta says has limited commercial terms that let most developers use it for free.
“Limited” is the word to watch. The coverage we found does not spell out the exact thresholds, so we can’t tell you where the line sits. If you plan to build a paid product on it, read the license text on the model page, not a summary.
The EU is a separate question. The Next Web notes the AI Act’s open-model exemption stops at systemic-risk models, and it does not say how Muse Spark Open is classified.
Can you run it on a Mac or a home PC?
Not on a normal one. An 816B model needs serious memory even with only 42B active. Meta showed it running on two Apple M5 Ultra Mac Studios using MLX, as Tech Insider reports.
That same report says independent checks of sustained speed are not available yet. Treat the demo as proof it can work, not proof it works well day to day.
If you want something smaller and local, Meta’s 30B Glimmer model, reported under Apache 2.0, is the more realistic pick. For another big open option, see Mistral Large 4.
How good is it, really?
Meta says it beats other open models, including Mistral Large 4, Qwen3.8, GLM 5.3 and Kimi K3, on agent and reasoning tests such as MCP-Atlas and HLE-Diamond. It trails Kimi K3 and GLM 5.3 on Terminal-Bench 2.1, a coding test.
Those are Meta’s own claims, and the coverage gave no scores. Wait for independent tests before you rely on “most capable US open-weight model,” which is how Wang described it on X.
Who should actually try Muse Spark Open?
Developers building coding assistants or multi-step agents are the obvious audience. A one-million-token window lets you feed in a whole codebase or a long set of documents without chopping it up first.
Companies with strict data rules may like the download option, since the weights can sit inside your own cloud account. Cloud marketplaces like SageMaker and Model Garden make that setup easier than it sounds.
If you only chat with AI now and then, skip it. A regular chatbot app will serve you better, and you will not need to think about licenses or servers.
Frequently Asked Questions
Is Muse Spark Open free to use?
The weights are free to download, and Meta says most developers can use it free under the Muse Model License. Hosted APIs charge their own prices.
Is Muse Spark Open the same as Muse Wonder?
No. Muse Wonder is Meta’s stronger, closed model. Muse Spark Open is the weaker one Meta chose to release.
How big is the Muse Spark Open context window?
One million tokens. It accepts text, images, video and audio, and replies in text.
Where can I try it fastest?
Use a hosted provider such as OpenRouter, Together AI or Groq. No downloads or hardware needed.
My take: try it through a hosted API this week, but hold off on building anything serious until the license terms and independent benchmarks are clear. It is a good sign that Meta is still shipping open weights. It is not yet proof this is the best open model you can pick.
Leave a Reply