--- title: "Meta's Muse Spark Open: open weights with terms. Fit for your product?" description: "The Muse Spark Open license lets you use, modify and sell the model for free, on condition of the phrase \"Built with Muse Spark Open\" and the use policy, with an agreement with Meta only for large API providers." url: https://ihsanalapsi.dev/articles/meta-muse-spark-open-licence lang: en date: 2026-10-21 updated: 2026-10-24 tags: ["AI", "Open models"] author: Ihsan Alapsi (إحسان العبسي) --- # Meta's Muse Spark Open: open weights with terms. Fit for your product? On October 15, 2026, Meta released the weights of Muse Spark Open: a model with 816 billion parameters, 42 billion of which run for each token, with a one-million-token context, built for coding and agent work. If you are thinking about building a product on it, the license decides what you are allowed to do, and the announcement sums it up as "limited commercial terms" that let many developers use it for free. The condition that touches every product is the phrase "Built with Muse Spark Open", while an agreement with Meta is only for those who sell access to the model as an API and whose company revenue is above 20 million dollars. ## Is Muse Spark Open open source? 1. Muse Spark Open is an open-weights model: you can download its weights, run them and modify them, but under Meta's own license, the Muse Model License, which comes with a use policy. The weights are the numbers the model learned during training, and they are the files you download to run it yourself. "Open source", on the other hand, has a specific definition: [the OSI's Open Source AI Definition](https://opensource.org/ai/open-source-ai-definition) (version 1.0) requires that you can use the system for any purpose without asking anyone for permission, and that the weights come with enough information about the training data and the full code used to build the system. Measured against this definition, the Muse license falls short on two points in its own text: the use policy forbids specific uses, and a large API provider must sign a separate agreement with Meta. And the model card describes the training data in a general paragraph: public sources, web crawling, data from third parties and from Meta products, and generated data. Meta itself calls it open-weight in the announcement and in the card. But the PyTorch team at Meta wrote in its post on the same day that the model is "a top open source choice for safety". So even inside the company the term is not fixed. Go back to the license text before you decide. ## What does the Muse Model License allow, and what are its conditions? The Muse Model License gives you, for free, the right to use, copy, modify (including fine-tuning), distill, merge, publish, distribute and sell. In return there are four conditions: a copyright and license notice with every copy, the phrase "Built with Muse Spark Open", an agreement for large API providers, and compliance with the use policy. The license version is dated October 15, 2026, and it has five sections followed by the use policy as an appendix (Exhibit A). The full text is on [the model page on Hugging Face](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B). | Section | What the text says | What it means for your product | | ------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------- | | 1 | The copyright notice, the license and the use policy must come with every copy of the model and of any Derivative, and your use must comply with the policy | If you distribute the weights or a modified version, the license and policy files go with them | | 2 | The phrase "Built with Muse Spark Open" must be displayed prominently in the interface of every product or service that uses it, or in its documentation, website or settings menu | Decide now where you will put it | | 3 | Anyone who gives others access to the model as a service (Model as a Service), and whose gross revenue together with its affiliates is above 20 million dollars in any 12 consecutive months, must notify Meta and sign a separate written agreement before offering the service | Applies to you only if you sell access to the model as an API and your company's revenue is above 20 million dollars | | 4 | Meta does not claim ownership of the outputs | Meta puts no condition on the text or code the model produces, but the use policy forbids outputs that infringe other people's rights | | 5 | The model is provided "AS IS" with no warranty of any kind, and you alone are responsible for the safety of any Derivative and how it is deployed | Testing and protection are on you | ### When does your own model count as a Derivative? The license defines a Derivative as any AI model "created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof". So the version you fine-tuned counts, and so does a small model you trained on data generated by Muse Spark Open. Distillation is exactly that: a small model learning from the answers of a bigger one. The card itself lists generating training data (Synthetic Data Generation) as one of the model's uses, so you could end up there without noticing. And if you do, the new model carries the phrase and a copy of the license, and the use policy too, because its text says it applies to the model "and any Derivatives". ## When do you need a separate agreement with Meta? You need a separate agreement with Meta in one case only: if you give others access to an AI model as a service, through an API or a similar interface, the other party receives its outputs, and the gross revenue of your company together with its affiliates is above 20 million dollars in any 12 consecutive months. The license calls this case Model as a Service. The figure is the revenue of the company and all its affiliates before expenses, not your revenue from the model alone, and not your profit. And the notice and the agreement must happen before you offer the model as part of that service. The same section has two exceptions: internal use, research and any commercial work that is not Model as a Service, and using the model as a component that powers features in your own product. | Your case | Does section 3 apply? | | ---------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------- | | An internal coding agent for your team | No, this is internal use | | A feature inside your app, such as summarizing reports or analyzing images for users | No, the model is powering a feature in your product | | An API you sell to other developers that returns the model's outputs, and your company's revenue is under 20 million dollars | Not now, but watch the number | | The same API, and your company's revenue has passed 20 million dollars in any 12 months | Yes: notify Meta and sign a written agreement before offering the service | Take the example of an online store that writes its product descriptions and answers customer questions through a chat built on Muse Spark Open. The chat is a feature inside the store, so it falls under the second exception, and it does not need an agreement with Meta however large the store's revenue gets. But the phrase "Built with Muse Spark Open" is still required, and the use policy forbids the chat from presenting its answers as if they came from a real employee. ## What does the use policy forbid? The use policy is part of the license, and the first thing in it is that Muse Spark Open is "not intended for individuals under the age of 18". After that comes a long list of forbidden uses, from violence and weapons to fraud and spam. Most of the list is what you would expect. These are the items that could touch a normal product: - Practicing a profession without a license, such as financial, legal, or medical and health advice. - Collecting or inferring sensitive information about people, such as health and identity, unless you have the legal right to it. - Writing malicious code or malware, or anything that disrupts a website or system. - Presenting the model's outputs as if they came from a person. - Failing to tell users about known risks of your AI system. The model card adds to the age rule: if your product could reach someone under 18, you are responsible for assessing and reducing the risks and for complying with the law. So if you are building an app for schools, factor in that responsibility from day one. For a health platform, the profession and sensitive information items need a reading with a lawyer before launch. The same goes for security tools and the malicious code item: it has no written exception for authorized testing, so a tool that generates an exploit for a penetration test needs the same reading before you build on it. ## How do you get Muse Spark Open, and how much hardware do you need? To run Muse Spark Open yourself, the model card asks for at least 700GB of GPU memory (VRAM) for the NVFP4 version, and 2TB for the BF16 version. That is why the realistic route for most people is a hosted provider. BF16 is the full weights in 16-bit numbers. The memory matches all 816 billion parameters, because the model is a Mixture of Experts: each layer has 256 "experts", 8 of which run for each token, but all 256 have to stay in memory. The NVFP4 version compresses each expert's weights to 4 bits and leaves the rest of the model uncompressed, and it is designed for NVIDIA Blackwell cards, but [Meta's developer post](https://dev.meta.ai/resources/blog/build-with-muse-spark-open) says it also runs on 8 H200 cards. | Route | What it needs | | ----------------------------------- | ------------------------------------------------------------------------------------------------------------- | | BF16 from Meta | 2TB VRAM per the card, and the files are about 1.6TB | | NVFP4 from Meta | At least 700GB VRAM per the card, the files are about 530GB, and it runs on 4 Blackwell Ultra cards or 8 H200 | | Mac Studio with MLX (NVFP4 version) | Two M5 Ultra machines with 512GB of unified memory each, or 4 machines with 256GB, that is 1TB in total | | 2-bit GGUF from Unsloth | About 300GB across RAM and VRAM in the post's text, and in its table a 256GB setup with llama.cpp | On top of the weights there is context memory (KV cache): according to [the Hugging Face blog](https://huggingface.co/blog/meta-muse-spark-open), a full one-million-token context takes about 266GB in BF16. And in a Hugging Face test on an M3 Ultra with 512GB, the 2-bit version in llama.cpp took 304GB and produced 18 tokens per second. The card warns that compressed versions can differ slightly in quality in rare cases. ### Requesting access on Hugging Face The original weights on Hugging Face are locked behind an access request, and approval is manual. The form asks for first name, last name, date of birth, country, organization and job title, with a warning above it: you must enter your "full legal name, date of birth, and full organization name with all corporate identifiers", you cannot edit the request after you send it, and a mistake can block you from this model and from others on Hugging Face. The information is stored and shared under Meta's privacy policy. There are other routes to the weights: Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. And Unsloth's compressed version on Hugging Face is not locked behind a request, but as the first section requires, it contains the same two files, `LICENSE.md` and `USE_POLICY.md`. ### Hosted providers Meta named seven API providers from day one: Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, OpenRouter and Together AI. That way you can try the model on your own work before spending a cent on hardware. The license is granted to anyone "accessing or obtaining a copy" of the model, and the use policy applies to "your access or use", so an API customer at a provider is most likely covered. And [the model page at Together AI](https://www.together.ai/models/muse-spark-open) points you to the same license and use policy. Treat Meta's terms as your terms, and read the provider's terms on top of them. ## Does it work well in Arabic, and how much can you rely on its numbers? It is not clear. The Muse Spark Open card says it was trained on data from more than 100 languages, but it does not name Arabic or publish the list of languages. It also says the model was not evaluated on all the languages in its training, and that performance can drop outside the well-supported languages. The same card asks developers to run safety and tuning tests for their own applications and "proposed languages". So if your product is in Arabic, the testing is on you: take a sample of real requests from your product, in Modern Standard Arabic and in the dialects your customers write in, and compare the answers with the model you run now, through a hosted provider. On performance, Meta compiled the card's table from different sources: numbers from [Artificial Analysis](https://artificialanalysis.ai/models/muse-spark-open) and official leaderboards such as Scale AI, numbers from Meta's own runs, and each company's numbers for its own model. [The methodology paper](https://research.meta.ai/static/muse-spark-open-evaluation-methodology) says their settings are not tuned specifically for the other models. And [TechCrunch](https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/) reported the performance as Meta's claim, "according to evaluation results released by Meta". The fairest comparison is the one where all the models were measured the same way. These are three rows from the card's table: the Artificial Analysis index (figures from October 7, 2026), Terminal-Bench 2.1 (all from Artificial Analysis with the same harness), and MCP-Atlas (Scale AI figures): | Measure | Muse Spark Open | GLM 5.3 | Kimi K3 | Qwen3.8 | | ------------------------- | --------------- | ------- | ------- | ------- | | Artificial Analysis index | 39 | 45 | 44 | 40 | | Terminal-Bench 2.1 | 76.4 | 83.9 | 85.0 | 82.0 | | MCP-Atlas | 89.2 | 84.2 | 82.3 | 84.5 | On the first two rows Muse Spark Open is behind all three. On MCP-Atlas, which measures tool use through real MCP servers, it comes first. Artificial Analysis gives it 39 on its page today, the same figure as in the table. And Meta's tagline, "the most capable US open-weight model to date", is about US models. In the same Meta table, GLM 5.3, Kimi K3 and Qwen3.8 score higher on the overall index. ## Where do you start if you want to build on it? Start with the license text, then decide your type of use, and try it through a hosted provider before any hardware. In order: 1. Open the model page on Hugging Face and read the license and the use policy in full. 2. Place your use in the section 3 table: a feature inside your product, or an API you sell to others. 3. Choose where the phrase "Built with Muse Spark Open" goes: the interface, the documentation, the website or the settings menu, and add it to your launch tasks. 4. Try the model through a hosted provider on a sample of Arabic requests from your product, and compare the answers with your current model. 5. If you will fine-tune it or train a model on its outputs, treat the result as a Derivative from day one, with the license and the phrase going with it. 6. Before you open it to the public, add the protections the card recommends: confirmation from a person before any action that cannot be undone, protection of the assistant's internal messages from user manipulation, and layers of protection against jailbreaks. If the model can call your API or read from your database, the permissions in your code decide how far it can go if someone manages to manipulate it. The [Security Skill Pack](/products/security-skill-pack) reviews your project's code and warns you about permission vulnerabilities before you deploy. After you add these protections, [Muhassan](/muhassan) tries to break them before launch, and hands you a clear fix plan. And if your daily work is with coding agents, there are ready-to-use skills for Claude Code and OpenCode in the [AI skills section](/products/c/ai-skills). Will your product use Muse Spark Open as a feature inside it, or will you sell access to it? ## FAQ **Does the phrase Built with Muse Spark Open have to appear in the app's interface itself?** Not necessarily. The license requires it to be shown prominently in the product's interface, or in the documentation, website or settings menu that comes with it. So a settings menu or a page on your site works, as long as it is visible. **Does Meta own the text and code the model produces?** No. Section 4 of the license says Meta does not claim ownership of the outputs. But if you train another model on those outputs, the new model counts as a Derivative, and the license and the use policy apply to it. **Can I fine-tune Muse Spark Open and sell the new model?** Yes, the license allows modifying and selling. But you must include the copyright notice, the license and the use policy with every copy, and write Built with Muse Spark Open. And if you sell it as an API service and your company's gross revenue has passed 20 million dollars in any 12 months, you need a written agreement with Meta first. **Is there a version of Muse Spark Open that runs on a normal laptop?** No. The smallest version Meta mentioned is a 2-bit GGUF version from Unsloth: Meta's developer post mentions about 300GB across RAM and VRAM, and elsewhere a 256GB setup. The realistic route for most people is a hosted provider. **If I use it through OpenRouter or Together AI, does the license apply to me?** Most likely yes. The license is granted to anyone accessing or obtaining a copy of the model, the use policy applies to your access or use, and the model page at Together AI points you to the same license. And read the provider's terms on top of it. **Can I use it in an app used by students under 18?** The text does not forbid it by name, but the use policy says the model is not intended for anyone under 18. And the model card makes you responsible for assessing and reducing the risks and complying with the law if minors could use it.