Muse Spark Open is a new model from Meta: 816 billion parameters in total, 42 billion active per token (a mixture-of-experts design), a 1M-token context window, and weights you can download from Hugging Face. Meta announced it on October 15 under a custom “Muse Model License” and calls it the most capable US open-weight model to date. That is Meta’s claim; I did not test it.
In this post I go over the Muse Spark Open license clause by clause, the usage rules that matter if you sell software to marketplace sellers, and what it costs through an API versus on your own hardware. My verdict: a startup can ship on it, the license is friendlier than I expected, but “open weights” and “open source” are two different things and this one is the first kind.
I am not a lawyer, and this is not legal advice. I read the text from the ggml-org mirror on Hugging Face, because the official Meta repo answered me with a 401. The mirror’s file is unchanged since I first saved it, but I cannot prove it equals the official one.
What does the Muse Spark Open license actually allow?
A lot. The grant says you may use, copy, modify (fine-tuning is named), distill, merge, publish, sublicense and sell, “without restriction”, subject to the conditions that follow. There are five numbered conditions, and for a normal startup only two of them bite.
Condition 2 is attribution. You must “prominently display ‘Built with Muse Spark Open'” in the user interface of each product that uses it, or in the docs, the website, or the settings menu. The settings menu is the easy out: one line of text.
The sneaky part is the word “Derivatives”. The license defines it as any AI model you create, train, modify, fine-tune or distill using the software “or any outputs or results thereof”. So if you generate training data with Muse Spark Open and use it to train your own small model, that small model is a Derivative. It carries the attribution line and the usage policy with it. Me, I would put it in the README of any model trained this way, on day one.
Condition 3 is the one everybody will quote, so here it is in plain words. If you sell access to AI models by API (Meta calls it “Model as a Service”), and you and your affiliates make more than 20 million US dollars in gross revenue over any twelve months, you must notify Meta and sign a separate agreement. The same condition says it does not apply to internal use, or to using the model “as a component to power features” of your own product that is not itself a model API. Trellis is nowhere near that line, so I did not lose sleep.
Condition 4 says Meta claims no ownership of outputs, and condition 5 is the usual “as is”.
Now the line to remember. Open source, as the Open Source Initiative defines it, cannot restrict fields of use. This license carries a usage policy and an attribution rule, so it is a custom license, and I will keep writing “open weights” for that reason.
Does the usage policy break anything for Amazon seller tools?
Not by itself, but a few lines read like they were written for our corner of the world. The policy is Exhibit A of the license and applies to every Derivative too. Section 3, on deception, bans:
- “Generating, promoting, or further distributing spam”
- “Representing that the use of Muse Spark Open or outputs are human generated”
- “Generating or facilitating false online engagement, including fake reviews and other means of fake online engagement”
Mettons you build a tool that drafts replies to customer reviews for sellers. A draft that a seller reads, edits and posts is normal assisted writing. A tool that writes 500 replies and posts them on its own, signed as the owner, walks right next to the “human generated” line. Anything that produces reviews, even “sample” ones, is over it. I made the scenario up to test the text, but every marketplace tool will meet its shape.
Section 1 also bans collecting or inferring “private or sensitive information about individuals” without the right to. Reviewer names and order data are personal data, so if your pipeline stuffs buyer details into prompts, talk to your lawyer.
My rule for Trellis would be a guard in front of every generation feature: a human approves before anything is posted, a log keeps what was generated, and reviews are a hard no.
What does Muse Spark Open cost: API or self-host?
Start with the API, because it is cheap. The OpenRouter model page lists it at $1.25 per million input tokens and $4.25 per million output tokens. They are OpenRouter’s prices, not Meta’s, and I did not compare other hosts.
Here is a worked bill, with a volume I made up. Say a review-reply feature writes 50,000 drafts a month. Each one sends 1,500 tokens in (the review, the listing, the seller’s tone notes) and gets 300 tokens back.
- Input: 50,000 × 1,500 = 75 million tokens × $1.25 = $93.75
- Output: 50,000 × 300 = 15 million tokens × $4.25 = $63.75
- Total: $157.50 a month
To cap a bill like that per customer, see my LiteLLM budget post. The short version: the cap needs a database behind it.
Now the other side. I added up the file sizes from the Hugging Face API for the GGUF repo (GGUF is the llama.cpp file format). Q2_K, the smallest, is 304.1 GB. Q4_K_S is 470.1 GB. Q8_0 is 871.6 GB. BF16 is 1,640.3 GB. These are my sums. “Self-host” does not mean a laptop, and not one big GPU either.
Meta’s own demo is two Apple M5 Ultra Mac Studios, which Meta says hold up to 512 GB each. I did not price that setup. I would guess it costs far more than $157.50 a month for a very long time, but I want a real quote before I name a break-even. Need the hardware for privacy? Fine. To save money, the API wins today.
A speed ceiling, and this one is inferred, not measured. Meta quotes 1.2 TB/s of memory bandwidth for the M5 Ultra. At Q4_K_S, the 42B active parameters are about 24 GB to read per token (470.1 GB × 42/816). 1,200 ÷ 24 gives roughly 50 tokens per second as a ceiling, on one machine’s bandwidth, before any overhead.
Is Muse Spark Open ready to run locally?
Closer than it was. The llama.cpp pull request, #30616, was merged on October 15 at 21:27 UTC. It was still open in my notes from the day before. The PR author says he tested it on a Mac Studio M3 Ultra and an RTX PRO 6000, and he found something I find useful: if the small latent projections are not kept at Q8_0, the low-bit quants turned the output into “a repeating loop”. The quantizer now keeps about 1% of the weights at Q8_0 to avoid it.
That is the author’s report. I did not download 304 GB to confirm it. I have no local run to show you. Merged tells me the architecture loads, not that Q2_K answers are any good for review replies, so I want other people’s results first.
What would I do on Monday?
Bon. Test it through OpenRouter, behind your own gateway, with a model name in config so you can swap it out. Put the “Built with Muse Spark Open” line in your settings page and the human-approval guard before anything posts. Check the license again before you ship, since Meta can publish a new version and I have only read the October 15 one.
Skip self-hosting for now unless you have the memory and a reason. The one thing I want to measure next is the review-reply task itself: the same 50 real-looking reviews through this model and through the one we use today, with a human scoring the drafts blind. If the cheaper model loses, the license will not matter.
