Muse Wonder costs $4 input and $20 output per million tokens, the same as Gemini 4 Argon from December 1; Muse Spark Open has 816 billion parameters and, even at 4 bits, needs about 408 GB for the weights alone, so running it locally takes a machine with more than 1 TB of RAM.
1. What Meta announced#
According to TechCrunch, on October 12, 2026 Meta Superintelligence Lab launched Muse Wonder and Muse Wonder Full, its first attempt at a frontier model, larger than Muse Spark, which was built for low latency. On October 15 it presented Muse Spark Open. The table summarizes what the two articles say.
| Model | Date | How to access it | Detail |
|---|---|---|---|
| Muse Wonder | Oct 12, 2026 | Meta Model API | $4 input and $20 output per million tokens |
| Muse Wonder Full | Oct 12, 2026 | Muse Code subscribers only; not in the API | No price published in the source |
| Muse Spark Open | Oct 15, 2026 | Unnamed API platforms; also local on Apple hardware | 816 billion parameters and a 1M-token context |
Meta compares Wonder's results with those of other leading models and highlights its «active memory»: retaining and synthesizing information the user provides. These are results from the company's own tests; verify them against your case before deciding.
Documentation: TechCrunch: Meta releases Muse Wonder (Oct 12, 2026) ↗ · TechCrunch: Meta launches an open-weight model that runs on Macs (Oct 15, 2026) ↗
2. What it costs against Gemini 4 Argon#
With the same example load (150 million input tokens and 30 million output tokens a month), Muse Wonder costs $1,200. Gemini 4 Argon costs $600 at the introductory rate until November 30 and $1,200 from December 1: the same list price as Muse Wonder. The peso figures use the TRM of October 17, 2026.
from decimal import Decimal as D
MODELS = { # USD per million tokens (input, output)
"Muse Wonder": (D("4"), D("20")),
"Gemini 4 Argon until Nov 30": (D("2"), D("10")),
"Gemini 4 Argon from Dec 1": (D("4"), D("20")),
}
TRM = D("3174.67") # TRM of Oct 17, 2026
INPUT_M, OUTPUT_M = D(150), D(30) # one month: 150 M input tokens and 30 M output tokens
for name, (t_in, t_out) in MODELS.items():
usd = INPUT_M * t_in + OUTPUT_M * t_out
print(f"{name}: ${usd:,.2f} a month | COP {usd * TRM:,.0f}")Price is only part of the decision. Meta's article does not state Muse Wonder's context window, and the quality results are the company's. Measure your task with both before comparing costs.
Documentation: Google: introducing Gemini 4 Argon ↗
3. Does Muse Spark Open fit your machine?#
The article says Muse Spark Open is designed to run locally on Apple computers, that it needs at least 1 TB of RAM (a figure TechCrunch attributes to Engadget) and that it performs best on a Mac Studio with the M5 Ultra chip. The code calculates how much memory the weights of 816 billion parameters take alone, by precision.
PARAMETERS = 816e9 # Muse Spark Open: 816 billion parameters
RAM_GB = 1000 # "at least 1 TB of RAM", per the source
for name, bits in [("16-bit", 16), ("8-bit", 8), ("4-bit", 4)]:
weights_gb = PARAMETERS * bits / 8 / 1e9
print(f"{name}: {weights_gb:,.0f} GB of weights only | fits in 1 TB: {weights_gb < RAM_GB}")The sources do not say what precision the weights are distributed in or how much memory a long context requires, so we cannot say how much RAM you need in practice. The 1 TB figure does indicate it is out of reach for a laptop or an ordinary desktop.
4. What is still unknown#
- Muse Spark Open's license: the article does not mention it, and it determines whether you can use it commercially.
- The names of the API platforms that offer it, and their prices.
- Muse Wonder's context window and model identifiers.
- Whether Muse Wonder Full will reach the API.
- Independent results: the benchmarks cited are Meta's.
5. How to decide whether to try them#
- For cloud production, compare Muse Wonder with the model you already use at the price that will apply, not the promotional one: use Codifly's AI model comparator.
- To run locally, confirm the license and the hardware first: without more than 1 TB of RAM it is not an option.
- If your team lacks that hardware, consider an API platform that hosts it once it publishes prices.
- Measure with your own tasks: a launch's results rarely match your case.
Go deeperHow to Choose an LLM for Production Without Relying on Leaderboards
6. Keep your numbers current#
This guide is based on press articles of October 12 and 15, 2026, not on Meta's technical documentation. Prices, licenses and availability may change; check Meta's official documentation when it publishes it.
- Confirm the price on the official Meta Model API page before budgeting.
- Check whether Meta publishes Muse Spark Open's license.
- Note the reading date of each rate.
Sources and scope
Documentation checked on October 18, 2026. Examples and decision criteria are editorial proposals; adapt them to your application's contract and validate them in an authorized test environment.