AI Tools
What does Muse Spark pricing look like: API or your own Macs?
Muse Spark pricing is $1.25 per million input tokens and $4.25 per million output tokens at Fireworks, Together AI and OpenRouter, checked on October 18, 2026. Meta publishes the 816B-parameter weights for free, but running them yourself needs two to four 512GB Mac Studios.
What matters
- Fireworks, Together AI and OpenRouter all list $1.25 per million input tokens and $4.25 per million output tokens; Fireworks and Together add cached input at $0.15.
- Fireworks calls it a limited preview through October 29, 2026, on US-only capacity, and does not support fine-tuning on that serverless route.
- Apple measured over 36 output tokens per second on four M5 Ultra Mac Studios at an 8,192-token context, and published no single-Mac number.
- The Muse Model License allows commercial use if you display 'Built with Muse Spark Open'; API resellers above $20 million in annual revenue need a separate Meta agreement.
- $10,000 of API spend buys about 2.35 billion output tokens at the $4.25 list rate.
What is Muse Spark Open?
Muse Spark Open is Meta's 816-billion-parameter open-weight model, released on October 15, 2026, with 42 billion active parameters per token and a 1-million-token context window. Meta describes it as a mixture-of-experts model built for coding, long-horizon agent work and multi-step reasoning. It accepts text, images, video and audio, and returns text only.
The weights live in a gated Hugging Face repository that requires manual approval, so access is a request rather than a download. Apple benchmarked the model in 4-bit NVFP4 precision, the variant a local deployment would realistically use, because the full-precision weights do not fit on desktop hardware.
Two things separate it from Meta's earlier Muse releases: scale (816B total parameters against 30B for Muse Glimmer) and delivery. Spark Open is weights plus hosted APIs. The Muse app people use on their phones is a different product with its own free tier.
What does Muse Spark Open cost through an API?
Three hosted routes publish the same list price as of October 18, 2026. Fireworks sells serverless access at $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens, with a 1,040k-token context. The model page tags it as a limited preview through October 29, 2026, served on US-only capacity with the usual 50% US-only premium waived, and lists fine-tuning as not supported.
Together AI shows the identical $1.25 / $0.15 / $4.25 and, at the time of writing, a limited-time free promo ending October 18, 2026. OpenRouter lists $1.25 input and $4.25 output with a 1,048,576-token context. Meta's launch post names Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together as API platforms; the other four were not checked for this article, so the table covers three verified routes, not the whole market.
Cached input is where recurring bills diverge. At $0.15 per million, cached tokens cost 88% less than fresh input, which matters for agent setups that resend a long system prompt on every turn. Loading a full 1,048,576-token context once costs about $1.31 at the input rate.
Can you run Muse Spark Open on your own Macs?
Apple's Machine Learning Research post from October 15, 2026 documents MLX inference distributed across Mac Studios over RDMA on Thunderbolt. Apple reports more than 36 output tokens per second on four Mac Studios at an 8,192-token context, using pipeline parallelism across two machines and tensor parallelism across two and four.
The floor is memory, not compute. Apple's Mac Studio specs page lists 512GB of unified memory as a configure-to-order option on the M5 Ultra with a 36-core CPU and 80-core GPU. Meta's post says two machines can hold an entire model: each M5 Ultra GPU addresses up to 512GB, and the 1.2TB/s memory bandwidth is what keeps decode moving.
Do the arithmetic before ordering anything. 816 billion parameters at 4 bits is roughly 410GB of weights before KV cache and runtime overhead, and Hugging Face lists 822 billion safetensors parameters, so a BF16 copy is about 1.64TB. That is why the usable configuration is two machines, not one.
What Apple did not publish is a single-machine number. Every figure in the post is a two- or four-Mac cluster, so for this model "runs on a Mac Studio" means a cluster of them. That distinction matters when someone budgets for one desktop and expects the same result.
What preview status means for a production decision
The cheapest-looking route is also the least permanent one. Fireworks ends its serverless preview on October 29, 2026, and Together's free promo is dated to October 18, 2026, so neither number should be hard-coded into a business case as a 12-month rate. Neither vendor publishes a rate that extends beyond its stated window, so the figures here are a snapshot, not a contract.
The practical move is to keep two providers configured rather than one, and to time any migration-heavy workload after the preview closes. If your product must be live in November, budget at the Fireworks list rate and treat the free and promo rates as a discount you might get, not a baseline you can rely on.
Capacity is the other variable. Fireworks serves this model on US-only capacity, which is fine for US teams and a compliance question for everyone else. Meta lists regional platforms such as Nebius and Baseten, but their rates were not verified here.
The break-even math: API versus a Mac cluster
At $4.25 per million output tokens, $10,000 of API spend buys roughly 2.35 billion output tokens. Hold that against the hardware bill.
Apple lists the 512GB configuration but does not publish its price on the specs page, and a two-machine setup also needs Thunderbolt 5 cabling, power and someone keeping MLX and the model weights current. If your workload will not push billions of output tokens through this model, the hosted rate is hard to beat on cash, and the preview window through October 29 is the procurement risk to plan around.
Local inference wins on a different axis. Data never leaves the machines, latency stops depending on someone else's queue, and the per-token cost stops tracking usage. Teams that already own 512GB Mac Studios and have a compliance reason to stay offline are the realistic local buyers. Everyone else should price a quarter of API usage before ordering hardware.
What the Muse Model License actually allows
The license is permissive on paper: use, modification, fine-tuning, distillation, redistribution and sale, with Meta claiming no ownership of outputs. Two conditions do the real work.
- Attribution: every product or service using the model or a derivative must display "Built with Muse Spark Open" in its interface, documentation, website or settings menu.
- Model-as-a-Service clause: if you resell model access through an API and your group's gross revenue passes $20 million in any 12-month period, you must notify Meta and sign a separate agreement before continuing.
Internal use, research, and embedding the model inside a product that is not itself a model service fall outside that clause. The Usage Policy in Exhibit A applies regardless, and the Hugging Face access gate is where you accept it.
How it compares with Meta's other Muse models
Muse Wonder lists at $4.00 input and $20.00 output per million tokens on OpenRouter, three to five times Spark Open's rate. It is a separate model, not a higher tier of the same one, so the price gap is not a like-for-like upgrade path. Muse Glimmer 30B sits at the other end at $0.30 input and $1.20 output with a 131,072-token context.
What the middle price buys is context and modality: a 1-million-token window and image, video and audio input on an 816B model. If your task is text summarization or short code edits, that capacity is idle spend.
Which route should you pick?
Start hosted. Fireworks or Together gives you the exact model at a published rate with no hardware commitment, and cached input at $0.15 per million is where long-prompt agent workloads save real money.
Move to local inference only when three things are true: you own or can justify two 512GB M5 Ultra Mac Studios, your data cannot leave your network, and your volume makes $4.25 per million output tokens the binding constraint. If any of those is false, the API is cheaper this quarter.
Drop to a smaller model if the task allows it. Muse Glimmer is a 30B model at $1.20 per million output tokens and fits machines that cannot hold Spark Open at all.
At a glance
| Route | Input / 1M | Cached input / 1M | Output / 1M | Catch |
|---|---|---|---|---|
| Fireworks serverless | $1.25 | $0.15 | $4.25 | Limited preview to Oct 29, 2026; US-only capacity; no serverless fine-tuning |
| Together AI | $1.25 | $0.15 | $4.25 | Limited-time free promo listed through Oct 18, 2026 |
| OpenRouter | $1.25 | Not listed | $4.25 | 1,048,576-token context |
| Self-host on 2 to 4 Mac Studios | No per-token fee | Not applicable | No per-token fee | Needs 512GB M5 Ultra machines; over 36 tokens/sec on four at 8K context |
| Muse Wonder | $4.00 | Not listed | $20.00 | Different Meta model, not the open-weight release |
| Muse Glimmer 30B | $0.30 | Not listed | $1.20 | Much smaller model, 131,072-token context |
FAQ
Is Muse Spark Open free to use?
The weights are free under the Muse Model License, including commercial use, as long as you display "Built with Muse Spark Open". Hosted inference is paid at $1.25 per million input and $4.25 per million output at Fireworks and Together, with a limited-time free promo at Together listed through October 18, 2026.
Can Muse Spark Open run on a single Mac Studio?
Meta's launch post says two Mac Studios can hold the entire model, and Apple's published benchmarks cover two- and four-machine clusters. A single Mac Studio tops out at 512GB of unified memory, and Apple did not publish a single-machine throughput figure.
How does Muse Spark pricing compare with Muse Wonder?
Spark Open lists at $1.25 input and $4.25 output per million tokens. Muse Wonder lists at $4.00 and $20.00 on OpenRouter. They are different models, so the gap is not a simple upgrade fee.
Related reading
Muse Glimmer local model, Meta Muse app pricing, cheapest LLM API, best local LLM