Meta released the weights for Muse Spark Open on October 15, 2026, and five hosts list it on OpenRouter at exactly the same price: $1.25 per million input tokens and $4.25 per million output tokens. There is no official Meta API to compare against. The hosts are the market, and for now they have all landed on one number.
That makes Muse Spark Open unusual among the open weight flagships we track. DeepSeek, Mistral, Z.ai and Moonshot each run their own API, which gives the market a reference price that third parties undercut or match. Meta sells nothing here. This page covers the price, the hosts, how it compares, and what the license means for who can sell it.
What Muse Spark Open is
From Meta's announcement and model card: Muse Spark Open is a natively multimodal mixture-of-experts model with 816 billion total parameters and 42 billion active, a 1M token context window, and text, image, video and audio input with text output. Meta calls it the most capable US open weight model to date and positions it for coding and agentic work.
Weights are on Hugging Face in BF16 and NVFP4. Meta says it runs on two Apple M5 Ultra Mac Studios, which is a first for a model this size and matters for the economics below.
Muse Spark Open pricing by host
OpenRouter listings as of October 20, 2026, USD per 1M tokens:
| Host | Input | Output | Cached input | Speed |
|---|---|---|---|---|
| Baseten | $1.25 | $4.25 | $0.15 | 175 tps |
| Nebius Token Factory | $1.25 | $4.25 | Not offered | 96 tps |
| Together | $1.25 | $4.25 | $0.15 | 70 tps |
| Modal | $1.25 | $4.25 | $0.15 | 69 tps |
| Fireworks | $1.25 | $4.25 | $0.15 | 54 tps |
Meta names seven launch partners: Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. Groq does not publish a Muse Spark Open rate on its public pricing page as of October 20, so it is not in the table.
Two things to read from it.
The spread is 1x. Five hosts, one price, on both input and output. We have seen this once before, on Kimi K3 in its first six weeks, when eleven hosts sat within 1.23x of each other. That market is now 45x wide on input. A single price at launch tells you where hosts started, not where they will stay.
Price is flat, speed is not. Baseten serves it at about three times the throughput of Fireworks at the same rate. When price does not separate hosts, speed, cache support and uptime do. Nebius is the only host of the five without a cached input rate.
What a real workload pays
A production assistant doing 10M input and 1M output tokens per day:
| Scenario | Daily | Annual |
|---|---|---|
| Any host, no caching | $16.75 | $6,114 |
| Any host with cache, 60% cache hits | $10.15 | $3,705 |
The cached rate of $0.15 is 12 percent of the input rate, so for prompt-heavy agent workloads the cache matters more than the choice of host. Nebius, without caching, is the expensive option for that kind of traffic even at the identical headline price.
How it compares
Large open weight models have settled into a narrow band on list price. USD per 1M tokens:
| Model | Input | Output | Source and date |
|---|---|---|---|
| Muse Spark Open | $1.25 | $4.25 | Hosts via OpenRouter, Oct 20 |
| Mistral Large 4 | $1.36 | $4.18 | Mistral, Oct 20 |
| DeepSeek V4 Pro, peak | $1.32 | $3.96 | DeepSeek, Oct 20 |
| DeepSeek V4 Pro, off-peak | $0.66 | $1.98 | DeepSeek, Oct 20 |
| GLM 5.3 | $1.40 | $4.40 | Mistral hosting, Oct 7 |
| Kimi K3 | $3.00 | $15.00 | Moonshot, Oct 8 |
Muse Spark Open, Mistral Large 4 and DeepSeek V4 Pro at peak are all within a few cents of $1.25 to $1.40 input and $4 to $4.25 output. Three labs, three continents, one price band. The only large open model priced well outside it is Kimi K3 at the top. DeepSeek's off-peak rate is the one way to pay half of it, and only on the official API during the off-peak hours.
For the full cross-model view, see the LLM API pricing comparison. For Mistral, see Mistral Large 4 API pricing. For DeepSeek, see DeepSeek V4 Pro API pricing.
The license decides who can sell it
Meta released the weights under the Muse Model License, which Meta describes as "limited commercial terms." The license text itself is gated behind a Hugging Face access request. Based on published summaries of it, the clauses that matter for buyers are an attribution requirement ("Built with Muse Spark Open" displayed in the product) and a revenue threshold: a provider offering the model as a service above roughly $20 million in annual revenue has to notify Meta and sign a separate agreement. Confirm both against the license text before building a commercial product on it.
For pricing, the threshold is the part to watch. It means the largest potential hosts, the hyperscalers and the big inference platforms, need a deal with Meta before they can list the model at scale. The five hosts on the table today are all under that umbrella one way or another. Whether more sellers show up, and whether the 1x spread holds, depends on how many of those agreements get signed. A license term is doing the job that an official API price does for DeepSeek or Mistral: it sets the shape of the market before a single token is sold.
What buyers should do
- Pick on speed and cache, not price. With one price across five hosts, throughput (54 to 175 tokens per second) and cached input support are the differences that show up on the bill and in latency.
- Budget at $1.25 and $4.25 and expect it to move. Every open weight market we have tracked has moved within two months of launch. A 1x spread is the starting point.
- Read the license before you depend on it. If your product is or could become a model service with meaningful revenue, the agreement clause applies to you, not just to the hosts.
- Lock what you can. When a price is this uniform, a forward contract fixes it before the hosts start competing in either direction. See how to lock in token prices and the Mercatus Forward Market.
To track token prices across models and hosts, see the Mercatus Token Index. Why one model carries one price here and twenty prices on Kimi K3 is covered in why token prices differ.
Frequently asked questions
How much does Muse Spark Open cost per million tokens?
$1.25 input and $4.25 output at every host listing it on OpenRouter as of October 20, 2026 (Baseten, Nebius, Together, Modal, Fireworks). Cached input is $0.15 where offered. Meta does not run its own API for the model.
Is Muse Spark Open free?
The weights are free to download under the Muse Model License for most uses, including fine-tuning and internal deployment. Running it yourself needs roughly 700GB of memory in NVFP4 or 2TB in BF16. API access is paid at the host rates above.
Who is the cheapest Muse Spark Open provider?
Nobody, yet. All five hosts charge $1.25 input and $4.25 output. Baseten is the fastest at about 175 tokens per second, and Nebius is the only one without a cached input rate.
Is Muse Spark Open open source?
Meta calls it open weight, not open source. The Muse Model License permits most uses but carries an attribution requirement and a revenue threshold for model service providers, per published summaries of the license.
How does Muse Spark Open compare to Mistral Large 4 and DeepSeek V4 Pro on price?
They sit within a few cents of each other at list: $1.25 and $4.25 for Muse Spark Open, $1.36 and $4.18 for Mistral Large 4, $1.32 and $3.96 for DeepSeek V4 Pro at peak. DeepSeek is half that off-peak on its own API.
Methodology
Model facts, release date, launch partners and license description from Meta's October 15, 2026 announcement and the Hugging Face model card. Host prices, cached input rates and throughput from OpenRouter's Muse Spark Open page on October 20, 2026. Groq checked on its public pricing page the same day and not listed. License clauses from published summaries; the license text is gated and was not read directly. Comparison rates from the vendor pages named in the table, on the dates shown. Workload figures use stated token volumes at 365 days; recompute with your own mix. Last verified: 2026-10-20.