Meta's Muse Spark Open Is Open-Weight, but the License Has Conditions. What to Read Before You Build on It

On October 15, Meta released the weights for Muse Spark Open. Meta calls it the most capable US open-weight model to date.

"Open-weight" is not the same as "no strings." The license has conditions that matter if you build a product on the model or train on its outputs. Here is what the sources say, and what I would read before committing.

What Meta says about the model

Per Meta's announcement:

• 816B total parameters, 42B active, natively multimodal, 1M-token context window.
• Built for coding and agentic work.
• Available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
• Weights available on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.
• Meta says it is the first in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with the latest MLX framework. It says the chip's unified memory lets two machines hold the entire model.

Meta also says the model met the thresholds of its Meta Superintelligence Scaling Framework for release, and that a full evaluation report is coming. The benchmark table in the post was not in the text I could read, so I am not quoting Meta's scores.

What an independent benchmarker lists

Artificial Analysis lists the model at 39 on its Intelligence Index v4.3.2 at xhigh reasoning, ranked #10 of 119 models. It reports about 102 output tokens per second, and prices of $1.25 per million input tokens and $4.25 per million output tokens, with a cost of $0.84 per Intelligence Index task.

Its summary calls the model particularly expensive compared with other open-weight models of similar size, though fast and fairly concise. The page also notes it supports text, image, speech and video input, with text output. Prices on hosted APIs may differ by provider, so check the provider you plan to use.

What the license says

The Hugging Face model page shows the Muse Model License, dated October 15, 2026. Meta describes it as having limited commercial terms that allow many developers to use the model for free. I read the text that page displayed. It started at condition 2, so I may have missed condition 1. Read the full text yourself.

From the conditions I could read:

• Attribution: you must prominently display "Built with Muse Spark Open" in the interface, documentation, website or settings menu of each product or service that uses the model or any Derivative.
• Derivatives are broad: the license defines them as any AI model created, trained, modified, fine-tuned, distilled, or otherwise improved using the model or any outputs or results thereof. Training another model on its outputs appears to be covered.
• Model as a Service threshold: if you or your affiliates give third parties access to AI models as a product via API, and your aggregate gross revenue exceeds $20 million over any consecutive twelve months, you must notify Meta and sign a separate written agreement before offering the model or Derivatives that way. The text says this does not apply to internal use, research, or using the model as a component of your own non-MaaS product.
• Outputs: Meta claims no ownership of outputs.
• Warranty: provided as is, and you are solely responsible for safety and deployment of your Derivatives.

The license also attaches a usage policy with prohibited uses, and says the model is not intended for people under 18. I did not review that policy line by line.

I am not a lawyer, and this is not legal advice. Have someone qualified read the license if your product depends on it.

What I would check first

These are suggestions, not tested results.

1. Read the full license, including condition 1. My reading came from a partial view.

2. Decide if you are Model as a Service. Selling API access to the model's outputs is different from using it inside your own product.

3. Check your training pipeline. If you distill or fine-tune on its outputs, find out whether that makes your model a Derivative with attribution duties.

4. Plan the attribution. Put "Built with Muse Spark Open" where the license requires, if you ship with it.

5. Compare hosted prices. Artificial Analysis called the model expensive for its size class, so compare providers before choosing.

6. Test on your own tasks. An index score is not your workload. I have not run the model.

7. Read the usage policy. It lists prohibited uses that may affect your product.

What I would not conclude

Do not read "open-weight" as "open source." Do not assume the license is the same as earlier Meta releases. And do not rely on this summary instead of the license text.

Frequently asked questions

What did Meta release?

The weights for Muse Spark Open, an 816B-parameter (42B active) multimodal model with a 1M-token context, on October 15, 2026.

Where can I get it?

Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted APIs include Together AI, Fireworks, Baseten, Modal, Groq, Nebius and OpenRouter.

Can it run on a desktop?

Meta says two Apple M5 Ultra Mac Studios can run it using MLX. I have not tested this.

Do I have to credit Meta?

The license text says to display "Built with Muse Spark Open" in products or services that use the model or its Derivatives.

Does training on its outputs count?

The license defines Derivatives as including models trained or improved using the model or its outputs.

When do I need a separate agreement?

For Model as a Service businesses with aggregate gross revenue over $20 million in any 12 months, per the text I read.

How does it score independently?

Artificial Analysis lists 39 on its Intelligence Index v4.3.2, ranked #10 of 119.

Sources

• Meta, Introducing Muse Spark Open: model details, availability, license summary (October 15, 2026).
• Hugging Face, Muse-Spark-Open-816B-A42B: Muse Model License text as displayed.
• Artificial Analysis, Muse Spark Open: index score, speed, pricing.
• The Rundown AI, October 16 issue: discovery pointer only.