Muse Spark Open for business: license, hardware and costs

On October 15, 2026, Meta published the weights of Muse Spark Open, which it calls “the most capable American open-weight model”. For anyone weighing Muse Spark Open for business, the license allows commercial use, including in the European Union, with two conditions: a “Built with Muse Spark Open” notice and a separate agreement for large companies that sell access to models through an API. Running it on your own server, though, means hardware with at least 700 GB of GPU memory, so for most companies the realistic route is a third-party API.

What Meta released

According to the model card on Hugging Face, Muse Spark Open is an MoE (mixture-of-experts) model:

Parameter	Value
Parameters	816B total, 42B active
Context window	1M tokens
Input	text, image, video, audio
Output	text
Intended use	coding, agents, long multi-step tasks
Weight versions	BF16 and NVFP4

Meta says the model was trained on data in more than 100 languages, but the same card notes that it was not evaluated on all of them. Polish is not named, so you have to check how well it handles Polish yourself.

In its comparison table, Meta puts the model next to other open models. It scores 39 on the Artificial Analysis index, alongside Mistral Large 4 (38) and behind GLM 5.3 (45) and Kimi K3 (44). It does lead on some agentic benchmarks, such as MCP-Atlas (89.2). These are Meta’s numbers; there are no independent tests yet.

Muse Spark Open license: what a company may do

The license is not Apache 2.0, but it isn’t restrictive either. It lets you use, modify, fine-tune, distill and sell solutions built on the model, under a few conditions:

1. Attribution. “Built with Muse Spark Open” must appear in the product interface, or in the documentation, on a website or in settings. This also applies to derivative models, including those trained on Muse Spark Open outputs.
2. Model as a Service above USD 20M. If a company, together with its affiliates, has more than USD 20M in revenue in any 12-month period and offers AI models to others through an API, it must sign a separate agreement with Meta before using Muse Spark Open in such a service. This condition does not cover internal use, or using the model as part of your own product that isn’t selling access to the model.
3. Acceptable use. The license comes with a list of prohibited uses. The ones that matter for companies include a ban on carrying out regulated activities without a license (legal, medical or financial, for example), a ban on passing off generated content as human-written, and a requirement to prevent use by people under 18.

One more thing matters for European companies: the license does not exclude the European Union. For licensees in Europe, the counterparty is Meta Platforms Ireland Limited.

How to run Muse Spark Open for business

There are three routes, from simplest to hardest:

1. A provider’s API. Meta lists Groq, Fireworks AI, Together AI, Nebius and OpenRouter, among others. On OpenRouter the model costs $1.25 per million input tokens and $4.25 per million output tokens. (check which region the provider processes data in if you send personal data)
2. Your own cloud instance. The weights are available in Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. This suits companies that want to keep data in a chosen cloud region.
3. Your own hardware. According to the model card, you need GPUs with at least 700 GB of memory for NVFP4 and 2 TB for BF16. Meta and Apple also showed a setup on two Mac Studios with M5 Ultra and 512 GB of memory each, but according to MacRumors that Mac Studio configuration isn’t on sale yet, and it’s more of a curiosity than infrastructure for a production deployment ;).

In my opinion, for a typical European company the weights themselves matter less than the fact that you can get the same model from several providers and, if needed, move it in-house. That reduces dependence on a single vendor, even if you use the API today.

What I would do now

1. Test the model on your own data through the API before you think about hardware. 50–100 real production tasks, including some in Polish, will tell you more than a benchmark table.
2. Watch the reasoning effort. The default is xhigh. Meta recommends low or medium when you want faster, shorter answers. That will most likely be enough for simple tasks, but check on your own data.
3. Add the “Built with Muse Spark Open” notice if you deploy the model for customers.
4. Check your AI Act obligations if you fine-tune the model and pass it on. I wrote about how the Commission is starting to enforce the AI Act.

If you’re looking for a replacement for models OpenAI is retiring, which I covered in my post on the GPT-4o API deprecation, Muse Spark Open is worth adding to the comparison.

TL;DR

Is Muse Spark Open free?

You can download the weights and use them without license fees, subject to the license terms. You pay for hardware or for a provider’s API.

Can Muse Spark Open be used commercially in the EU?

Yes. The license does not exclude the European Union. A separate agreement with Meta is only needed for companies with more than USD 20M in revenue that sell access to models through an API.

How is Muse Spark Open different from Muse Glimmer?

Muse Glimmer is a smaller model (30B parameters), released in August under the Apache 2.0 license, according to VentureBeat. Muse Spark Open is much larger and has Meta’s own license.

Does Muse Spark Open work in Polish?

Meta claims training on more than 100 languages but doesn’t publish a list of supported languages. Test it on Polish text before you deploy it.