Muse Spark Open: Downloadable Weights, Hardware Needs and License Terms

Meta released Muse Spark Open on October 15, 2026. Developers can download the weights or use a hosted provider, giving teams another way to run a large coding and tool-using model outside Meta's own applications.

The practical decision is where to run it. An 816-billion-parameter model needs a serious deployment plan even when it activates only part of the network for each token.

What Muse Spark Open includes

The official model card specifies 816B total parameters, 42B active parameters and a one-million-token context window. It accepts text, images, audio and video, and produces text.

Meta positions it for coding, multi-step reasoning and tool use. Its safety post-training is intended for deployment outside Meta's hosted guardrails. The model card still recommends adding safeguards suited to the application and testing the model in the languages and workflows a team expects to support.

That matters for an agent with repository or business-system access. Choosing a model is only one part of deciding which files it can read, which tools it can call and when someone must approve its next action.

Downloads and hosted access

Meta's launch announcement lists immediate hosted access through providers including Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI. Weights are distributed through Hugging Face and major cloud model catalogs. Hugging Face requires an account and review of the access conditions.

The release supports inference stacks including vLLM, SGLang, TensorRT-LLM, Transformers, MLX and llama.cpp. Existing infrastructure will help determine which route is worth testing first.

For GPU deployment, the model card lists at least 700GB of VRAM for NVFP4 and 2TB for BF16. Weight format is therefore a central part of the hardware decision.

Meta also demonstrates running the model on two M5 Ultra Mac Studios, each with 512GB of unified memory. That is a specific two-machine configuration. The announcement does not establish performance on an ordinary laptop, or that every Mac Studio configuration can hold the model.

For a team comparing hosted and local use, the useful measurements are accepted task completions, peak memory and total operating cost. Keeping weights in-house adds responsibility for updates, monitoring and access control.

Muse Spark Open on Mac Studio: measured performance

Update, October 16, 2026: Apple's October 15 MLX benchmarks add deployment measurements for Muse Spark Open in NVFP4. With an 8,192-token input context and batch size one, Apple measured decode speed over 128 generated tokens:

M5 Ultra Mac Studio configuration	Decode tokens per second	Time to first token
Two machines, pipeline parallelism	15.74	8.4 seconds
Two machines, tensor parallelism	24.99	8.2 seconds
Four machines, tensor parallelism	36.73	4.8 seconds

The setup uses 512GB Mac Studios and RDMA over Thunderbolt 5, with two cables per peer-to-peer link. Pipeline parallelism puts different layers on different machines; tensor parallelism splits computation within layers.

These are Apple's measurements, not independent tests or an application-level latency guarantee. They also do not establish performance at the model's full one-million-token context. At 65,536 tokens, Apple reports that two-machine tensor parallelism takes longer to produce the first token than two-machine pipeline parallelism. Teams should measure their own prompt lengths and concurrency before choosing the more complex setup.

Read the Muse Model License

The published license includes attribution and usage-policy requirements. Products using the model or its derivatives must display the prescribed Muse attribution in an allowed location.

A separate condition covers businesses providing third parties with model outputs through an API or similar interface. If the licensee or an affiliate offers such a model-as-a-service business and their combined gross revenue exceeds US$20 million in any consecutive twelve months, the licensee must notify Meta and obtain a separate written agreement before offering Muse or its derivatives through that business. The revenue test covers the licensee and affiliates collectively. Exceptions include internal use, research and commercial products that are not model-as-a-service offerings.

Those conditions belong in a deployment review alongside infrastructure and evaluation results. Downloadable weights give a team more control over its system, while leaving it responsible for how that system behaves.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn complex product launches into clear stories for developers and customers.

Book a 30-minute call.