Muse Spark Open tested: Meta's open model, its licence and Muse Wonder

Meta has a new AI model you can download: Muse Spark Open. Meta calls it "the most capable US open-weight model to date". Three days earlier it launched Muse Wonder, which Meta calls its first "frontier" model. That one you can only use online. We ran both on our own secret coding tests. Both got every check right. But Muse Spark Open was slow and wordy, so one job cost us two to four times as much as with Claude Sonnet 5.5 or GPT-6.1 Sol. And "open" comes with rules.

What did Meta release?

Two things in one week:

• Muse Wonder, on 12 October. Meta's Superintelligence Lab calls it Meta's first "frontier" model, a model meant to compete with the very best. You use it through an API, the paid connection for programs, at $4 per million tokens in and $20 out, TechCrunch reports. Meta says it will also power its Muse agent. A bigger version, Wonder Full, is only for subscribers of Muse Code, Meta's coding tool.
• Muse Spark Open, on 15 October. This one has open weights: the files that make up the model, which anyone can download from Hugging Face. It has 816 billion parameters, of which 42 billion work on each word, and it can read up to a million tokens at once. Tokens are the small pieces of text an AI reads and writes.

Meta already had Muse Spark models before, such as Muse Spark 1.3. Those only work through an API; we found no download for them. We tested 1.3 too, to see what changed.

We wrote about Meta's Muse app, the helper that works for you in the background, in an earlier article. This article is about the models.

Is Muse Spark Open really open?

Yes, but with Meta's own rules. The Muse Model License starts like a very free licence: you may use, change and share the model "free of charge". Then come two rules:

1. Show the name. Any product built with it must "prominently display" the words "Built with Muse Spark Open", in the app, the documentation, the website or a settings menu. 2. Big AI-model services need a deal. Say your company, counted together with its sister companies, earns over 20 million dollars in any 12 months, and it lets others use AI models through an API or something similar. Then it must tell Meta and sign a separate agreement before it offers Muse Spark Open in that service. This rule does not apply to research, to use inside your own company, or to an app of your own that uses the model, as long as that app does not itself give others access to models.

Two more details. A model that you train with Muse Spark Open's answers counts as a "derivative" under the licence, so the same rules apply to it. And Meta "claims no ownership" of the answers the model gives you.

To download the files, Hugging Face asks you to accept the licence and fill in your full name, date of birth and company details. We did not do that, because we tested the model through an online service.

Can you run it on your own computer?

Not a normal one. Meta's model card asks for at least 700 GB of graphics memory for the smaller version of the files, and at least 2 TB for the full version. Meta says it runs on two Mac Studio computers with Apple's M5 Ultra chip, work it did together with Apple, TechCrunch reports.

So most people will use it online. On OpenRouter, a service that gives access to many models, it costs $1.25 per million tokens in and $4.25 out. Five companies offered it there when we looked, from Baseten to Together. That is the same price as the older Muse Spark 1.3.

Our test: three small coding jobs

We used our private secret test set. Our older test is public, so new models may have learned its answers. This one has three jobs:

• D. Fix a money reader. Code that turns Dutch amounts like "€ 1.234,56" into cents has several bugs. 28 checks.
• E. Split a restaurant bill. From a one-page description, write code that splits a bill. Shared dishes are split evenly, a tip is added, and the cents must add up exactly. 17 checks.
• F. Read a CSV file, a spreadsheet saved as text, where commas and quotes can hide inside a field. 17 checks.

Secret checks are tests we wrote ourselves and only run after the AI has answered. Two examples: "€ 0,5" must give 50 cents, and a €10 bottle of wine for three people must come to 334, 333 and 333 cents.

Each model did each job three times, in one request through OpenRouter (GPT-6.1 Sol: through OpenAI itself), with the default settings. That is 186 checks per model. We compare the results with models we tested on the same jobs between 7 and 10 October.

The results: every check right, but slow

Muse Spark Open got every check right, three times out of three. Among the open models we ran on all three jobs, it is the only one with a perfect score: GLM-5.3 and DeepSeek V4 Pro each missed a few checks. Muse Spark 1.3 missed one, in the CSV job.

But it was not quick. A typical job took about 93 seconds, against 17 for Claude Sonnet 5.5. And it cost about 5.0 cents per job, about 2 times as much as Sonnet and 4 times as much as GPT-6.1 Sol. Our price overview has more models.

Why it cost more than its price suggests

Per token, Muse Spark Open is cheaper than Sonnet 5.5 or GPT-6.1 Sol. So why did a job cost more?

Because it thinks out loud for a long time. In the middle of our runs, it wrote about 12,700 tokens per answer. Sonnet wrote about 2,400, and Mistral Large 4 about 1,200. You pay for every token, also for the thinking you never see.

Muse Spark 1.3 did the same, with about 9,400 tokens. Meta's model card says you can turn the thinking down, with a setting called reasoning effort. We used the default settings, as we do for every model.

Muse Wonder: faster, but pricier

Muse Wonder also got all 186 checks right. It was much faster than Spark Open, about 31 seconds per job, and wrote far fewer tokens. But per token it costs three to five times more ($4 and $20 on OpenRouter), so one job came to about 8.0 cents. That is more than three times the cost of Claude Sonnet 5.5. Only GLM-5.3 cost more in our table.

Meta says Wonder scores close to Claude Fable 5.1 and GPT-6 Astra on its own tests, TechCrunch reports. Our three small jobs are too easy to check that: Sonnet 5.5 and GPT-6.1 Sol also got everything right.

Meta says Wonder will also power its Muse app, the agent we wrote about before. Separately, in September, before Wonder came out, a columnist at Inc. claimed that the Muse app had read his private messages on his Mac without permission. Meta's spokesperson Andy Stone said the Messages link "is entirely opt-in", TechCrunch reported. The two sides still disagree. This is about the Muse app, not about using the models through an API.

What Meta's own scores say

Meta's model card compares Muse Spark Open with other open models, and that table is worth a look. It shows Chinese models ahead on several tests:

• On the Artificial Analysis Intelligence Index, a combined score, Spark Open gets 39. GLM 5.3 gets 45 and Kimi K3 gets 44.
• On Terminal-Bench 2.1, a coding test, it gets 76.4. GLM 5.3 gets 83.9 and Kimi K3 85.0.
• Spark Open is best in Meta's table on MCP-Atlas (using tools), JobBench (work tasks) and a set of very hard questions without tools.

So "the most capable US open-weight model" is a careful choice of words. Meta's own table shows Spark Open ahead on some tests and behind on others. It does not show a clear lead over all open models.

The table also has a safety test, StrongREJECT, where lower is better: Spark Open scored 0.5, the others about 63 to 86. These are all Meta's own numbers, on tests Meta chose.

Should you use it?

It depends on what you want:

• You want open weights you may build into a product: Muse Spark Open is a strong choice, and it was right every time in our test. Remember the "Built with Muse Spark Open" label.
• You pay per job through an online service: it cost us more than Sonnet 5.5 or GPT-6.1 Sol for the same result, and it was slower. Try a lower reasoning effort, or pick another model.
• You want to run it at home: not realistic, unless you have hundreds of gigabytes of graphics memory.
• You want Meta's top model: Muse Wonder was right every time and three times faster than Spark Open, but the most expensive of the Meta models.

These were small, clear jobs. They do not show how the models do on long projects.

How we checked

On 17 October 2026 we read Meta's blog post, the model card and the Muse Model License on Hugging Face, OpenRouter's pages for Muse Spark Open and Muse Wonder, and TechCrunch's reports of 12 October, 15 October and 30 September. The scores in "What Meta's own scores say" come from Meta's model card; we did not run those tests.

The same day we ran Muse Spark Open, Muse Wonder and Muse Spark 1.3 on our private secret test set through OpenRouter: three jobs, three times each, 186 checks per model, with the default settings. There were no errors. Time, tokens and cost come from OpenRouter's reply. The three Meta models cost us about $1.56 together, paid by the owner of this site. The other models in the tables ran on the same jobs between 7 and 10 October 2026 (GPT-6.1 Sol through OpenAI itself).

We did not download the model files and did not create a Meta account. The screenshots are our own, logged out and cropped. The charts were made by us, and none of the images was made by AI. Models, prices and licences can change.

Correction, 19 October 2026: we first wrote that Muse Wonder powers Meta's Muse agent. TechCrunch wrote that it "will power" the agent, so we changed that in three places.