How Much To Run AI
Open weights

Muse Spark Open: API costs and multi-device setup

Muse Spark Open is Meta’s open-weight model for coding, multimodal reasoning and tool agents, released October 15, 2026. It has 816B total parameters, 42B active per token and a 1M-token model context. Hosted inference is the practical first step for most users; self-hosting needs substantial GPU or unified memory, with hardware requirements depending on the quantization and runtime.

Content reviewed:

Muse Spark Open has 816B total and 42B active parameters; a Together request with 100K uncached input and 10K output costs $0.1675.
Budget and deployment illustration. The illustrated NVFP4 route uses multiple configured Macs.

Use this model

API docs and console

Official model repository

Local tools and guides

Use cases

  • Use repository tools to inspect, edit and test a bounded coding change.
  • Ask a text question about images, audio or video through a backend that supports those inputs.
  • Run a private agent service when you can operate the full model and its tool environment.

Try Muse Spark Open through hosted inference

Meta lists Together AI, Fireworks, Baseten, Groq, Modal, Nebius and OpenRouter as inference partners. Pick a provider, create its API key and select that provider’s exact Spark Open model ID. Together’s serverless route uses meta-models/Muse-Spark-Open-816B-A42B. For the Python example, install the together package and set TOGETHER_API_KEY in your environment.

The model accepts text, images, audio and video and returns text. Individual backends can expose fewer input types or a smaller context than the checkpoint. For a coding agent, connect the model to your repository tools and tests; the model’s response alone does not execute a file edit.

Together SDK example; running it incurs your account’s API charges
from together import Together

client = Together()
response = client.chat.completions.create(
    model="meta-models/Muse-Spark-Open-816B-A42B",
    messages=[{"role": "user", "content": "Plan a repair for an invoice parser that accepts blank IDs. Include edge cases."}],
)
print(response.choices[0].message.content)
print(response.usage)

Together API pricing and a request budget

Together’s published USD rates, checked October 16, are $1.25 per million uncached input tokens, $0.15 for cached input and $4.25 for output. Promotional access credits or a temporarily free route are separate from these ordinary usage rates.

For 100,000 uncached input tokens and 10,000 output tokens, input is $0.125 and output is $0.0425: $0.1675 total. If all of that input is billed as cache hits, the same output brings the total to $0.0575. Count the billed output across reasoning, tool turns and retries when budgeting a complete agent task.

Together serverless usageUSD / 1M tokens
Uncached input$1.25
Cache-hit input$0.15
Output$4.25

Choose a checkpoint and a supported server

All 816B parameters must be stored and available to the runtime even though a token activates 42B. Meta’s current download guide lists about 537GB of NVFP4 weights or 1.65TB in BF16. Its serving recipes use four B300 GPUs for NVFP4 or eight for BF16. Weight footprints exclude context caches, activations and runtime allocations.

Use the runtime recipe for your accelerator and input types. Meta maintains SGLang, vLLM, TensorRT-LLM, Transformers and other deployment paths; the platform guide also covers AMD and Intel. Download access on Hugging Face requires accepting the model conditions. After loading, check the server’s model list and send a short request before raising context or concurrency.

Run across two or four M5 Ultra Mac Studios

Meta’s Apple configuration uses NVFP4 with two 512GB or four 256GB M5 Ultra Mac Studios. MLX distributes the model across the machines; RDMA over Thunderbolt 5 provides the interconnect. Apple reports a roughly 493GB model footprint in its MLX test before intermediate computations, so that NVFP4 configuration uses multiple machines.

Apple measures 24.99 decoded tokens per second on two machines with tensor parallelism and 36.73 on four, using an 8,192-token prompt and 128 generated tokens. These are publisher measurements at that prompt size. Plan long-context memory and prefill separately from those decode figures.

Price the exact memory configuration, storage and interconnect for every node. A starting Mac Studio price does not quote a 512GB machine or the complete cluster. The hardware specifications and current order availability belong in that purchasing check.

MLX server entry point after configuring all hosts and the RDMA link
mlx.launch --hostfile hostfile.json --backend jaccl --env MLX_METAL_FAST_SYNCH=1 -- \
  mlx_lm.server --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --port 8080

A different quantization fits one 512GB M3 Ultra

Hugging Face’s integration team tested the oQ3e quantization on one M3 Ultra with 512GB unified memory. It reports about 344GB occupied by the model, 225 tokens per second in prefill and 21 tokens per second in generation with a 1,024-token prompt and 128 generated tokens. This is an ecosystem team’s short-context test with a different checkpoint, not an independent user reproduction of Apple’s multi-machine NVFP4 run.

The model footprint is not the peak memory of a serving workload. Budget KV cache and runtime allocations for your context and concurrency; this result does not demonstrate 1M-token or multi-user serving on one Mac.

How the Muse Model License treats commercial use

The Muse Model License permits commercial uses subject to its terms, attribution and usage policy. A separate agreement is required when you or an affiliate offer model-as-a-service access and combined gross revenue exceeds US$20 million in a consecutive twelve-month period. That clause expressly excludes internal use and commercial products that use the model as a component without being model-as-a-service businesses.

Keep the required Built with Muse Spark Open attribution and license notices when applicable. Offering a model API is a different licensing question from using Spark Open internally to develop an application.

Compare Spark Open with the service you need

Muse Wonder and Wonder Full have separate hosted access and subscription arrangements. Choose Spark Open when downloadable weights or a supported self-hosted deployment matter. Compare a completed coding task with the same tools and acceptance tests, rather than applying Wonder’s prices or benchmark scores to Spark Open.

Nemotron 3.5 offers a smaller open-weight vision-language family; LFM2.5 is a text-only local-agent option. Compare their actual hardware, input types and license terms before using parameter counts to choose a runner.

Before you use it

  • Input formats, context and serving capacity depend on the selected hosted provider or self-hosted backend.

How to get started

Try one bounded task through a hosted provider, retain the total token usage, and evaluate self-hosting against a quoted, fully configured multi-device system.

Sources

Sources (9)

Calculators