Muse Spark Open: Self-Host vs Hosted Cost (2026)
Updated October 16, 2026
The Muse Spark Open weights are now public. This page compares the two ways to run it: self-host it on your own hardware, or pay hosted providers per token. Every figure below links to an official or provider page. Anything we could not confirm is marked Not verified.
What was released
- One model: Muse Spark Open, 816B total parameters, 42B active, sparse mixture of experts (official docs).
- Multimodal input: text, image, audio and video in, text out. Context window of 1,048,576 tokens (official docs).
- Checkpoints: BF16 (about 1.65 TB) and NVFP4 (about 537 GB) (official docs).
- License: Muse Model License with limited commercial terms, plus an Acceptable Use Policy (official announcement, Oct 15, 2026).
- Runtimes: SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth (announcement).
Self-host vs hosted cost comparison
Prices checked: October 16, 2026
| Option | Hardware / VRAM | Price | Rough monthly cost | Source |
|---|---|---|---|---|
| Self-host, BF16 | ~1.65 TB of weights. Budget ~2 TB of VRAM. 8 GPUs, e.g. 8× B300 | Hardware/rental | Not verified | Official docs |
| Self-host, NVFP4 | ~537 GB of weights. One 4-GPU NVIDIA Blackwell node | Hardware/rental | Not verified | Official docs |
| Self-host, desktop (MLX) | 2× Apple M5 Ultra Mac Studio, up to 512 GB unified memory each | One-time purchase | Not verified | Announcement |
| Together AI | None (serverless) | $1.25 in / $4.25 out / $0.15 cached input, per 1M tokens | Usage-based | Together AI announcement |
| Fireworks AI | None (serverless) | $1.25 in / $4.25 out / $0.15 cached input, per 1M (serverless limited preview through Oct 29, 2026) | Usage-based | Fireworks model page |
| Baseten | None (serverless) | $1.25 in / $4.25 out / $0.15 cache input, per 1M | Usage-based | Baseten model page |
| Modal | None (serverless) | Not verified Listed price from OpenRouter: $1.25 / $4.25 | Usage-based | Modal shows token rates only in its dashboard (Modal docs). OpenRouter price: OpenRouter model page |
| Nebius Token Factory | None (serverless) | Not verified Listed price from OpenRouter: $1.25 / $4.25 | Usage-based | Nebius prices page did not load for us. OpenRouter price: OpenRouter model page |
| Groq | None (serverless) | Not verified | Usage-based | Groq pricing did not show a Muse Spark Open price. Groq is not listed on the OpenRouter model page. |
We did not find an official GPU rental price for B300 or Blackwell nodes running this model, so we do not estimate a self-host monthly cost. Example usage math: 10M input + 2M output tokens at the $1.25 / $4.25 list rates confirmed on the Together, Fireworks and Baseten pages = 10 × $1.25 + 2 × $4.25 = $21.00 before caching.
Who should choose which
Choose hosted if…
- Your monthly usage is low or bursty and you would rather not keep 2 TB of VRAM idle.
- You want to start today. Together, Fireworks and Baseten publish the same list price.
- You do not have a team to run a multi-GPU serving setup.
Looking for a cheaper hosted option? See DeepSeek V4.1 Flash API pricing.
Choose self-host if…
- Prompts and responses must stay inside your own infrastructure. The official docs call this out as the main reason to self-host.
- You already own or have reserved Blackwell-class GPUs. With NVFP4 the model fits on one 4-GPU node.
- You plan to fine-tune with PyTorch/TorchTitan, or you need full control of the runtime.
Choose the desktop (Mac Studio) route if…
- You are one developer or a small team experimenting locally and can buy two M5 Ultra Mac Studios.
FAQ
What is Muse Spark Open?
The open-weight release of Muse Spark, released October 15, 2026: a sparse mixture-of-experts multimodal model with 816B total and 42B active parameters and a 1,048,576-token context window, under the Muse Model License.
How much VRAM does Muse Spark Open need to self-host?
Per the official docs, the BF16 checkpoint is about 1.65 TB of weights, serves on eight GPUs such as 8× B300, and needs roughly 2 TB of VRAM in total. The NVFP4 checkpoint is about 537 GB and serves on a single 4-GPU NVIDIA Blackwell node.
How much does hosted Muse Spark Open cost per 1M tokens?
As of October 16, 2026, Together AI lists $1.25 per 1M input tokens, $4.25 per 1M output tokens and $0.15 per 1M cached input tokens. Fireworks and Baseten list the same rates on their own model pages. We could not confirm Modal, Nebius or Groq on their own pages.
Can Muse Spark Open run on a Mac?
The official announcement says it runs on two Apple M5 Ultra Mac Studios using MLX, with up to 512 GB of unified memory per machine.
Is Muse Spark Open free for commercial use?
It is released under the Muse Model License with limited commercial terms, plus an Acceptable Use Policy. Read the license itself before using it commercially.