How to Use Muse Spark Open: Download, Run and API Guide (2026)

Muse Spark Open is Meta’s new open-weight AI model, released on 15 October 2026, and Meta describes it as the most capable US open-weight model to date. Unlike the Muse Spark model inside the Meta AI app, you can download the weights, run them on your own hardware or call them through third-party APIs. This guide explains what Muse Spark Open is, what the licence allows, the hardware you realistically need, and four practical ways to start using it today, from a two-minute API call to a full self-hosted deployment.

Last updated: 16 October 2026. Details below were checked against Meta’s announcement, the official Hugging Face model card and licence, and OpenRouter’s model page on 16 October 2026. We have not run the 816B model ourselves; commands come from Meta’s documentation, and performance figures are Meta’s own unless stated otherwise.

What is Muse Spark Open?

Muse Spark Open is a large mixture-of-experts (MoE) language model from Meta Superintelligence Labs. It is part of the Muse family that also includes the hosted Muse Spark assistant model (covered in our guide to Meta Muse) and the smaller Muse Glimmer 30B. The headline specifications are:

Because only 42 billion parameters are active for each token, inference is far cheaper than the 816B total suggests. The catch is memory: all 816 billion parameters still have to be loaded somewhere, which is why hardware is the first thing to think about.

Is Muse Spark Open open source?

Not in the strict sense. Muse Spark Open is an open-weight model released under the Muse Model License, not under an OSI-approved licence such as Apache 2.0 or MIT. You can download, run, fine-tune and build products on the weights, but the licence adds conditions. The main ones are:

• Large companies must talk to Meta first. If your organisation earned more than $20 million in revenue over the previous 12 months, you must notify Meta and sign a separate agreement before offering the model as a model-as-a-service product.
• Attribution is required. Products built on the model must display “Built with Muse Spark Open”.
• Some uses are banned. The licence prohibits military and critical-infrastructure uses, with a carve-out for the US government.

For most individual developers, researchers and small businesses, the licence is workable. Larger organisations, and anyone in defence, energy or similar sectors, should read the full Muse Model License with their legal team before building on it. If you need a genuinely permissive licence, Muse Glimmer 30B (covered below) uses Apache 2.0.

Muse Spark Open hardware requirements

This is where most people’s plans change. According to the official model card, you need roughly:

• About 700 GB of GPU memory for the NVFP4 weights.
• Roughly 2 TB of GPU memory for the full BF16 weights.

In practice that means a multi-GPU data-centre server, a cloud instance with several high-memory accelerators, or Meta’s demonstration setup: two Apple M5 Ultra Mac Studios with 512 GB of unified memory each, linked together and running Apple’s MLX framework. Note that, according to VentureBeat’s launch report, the 512 GB Mac Studio configuration was not yet on sale at launch and is expected in late October.

No laptop, including the new NVIDIA-powered machines in our RTX Spark laptop guide, can hold the full model. Those machines are aimed at models up to about 120B parameters. If you want something you can run locally on one machine, skip to the Muse Glimmer section.

How to use Muse Spark Open: four options

Pick the route that matches your hardware and budget. Most readers should start with Option 1 and only move to self-hosting once they know the model suits their work.

Option 1: Use Muse Spark Open through OpenRouter (fastest)

OpenRouter lists the model as meta/muse-spark-open. At the time of writing its listed price was $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15 per million, and the full 1,048,576-token context. Prices are in US dollars and can change, so check the OpenRouter model page before you budget.

1. Create an OpenRouter account and add credit.
2. Generate an API key in your account settings.
3. Send a request to the OpenAI-compatible chat completions endpoint with the model name set to meta/muse-spark-open.

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta/muse-spark-open",
    "messages": [{"role": "user", "content": "Summarise this contract clause in plain English: ..."}]
  }'

Meta also lists Baseten, Fireworks, Groq, Modal, Nebius and Together as API partners, so you can compare prices and data-handling terms across several providers. If you already use other model APIs, our DeepSeek API guide shows the same OpenAI-compatible pattern in more detail.

Option 2: Deploy it from a cloud model catalogue

Meta says the model is available in Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. This route suits companies that already have a cloud contract, need data to stay in a particular region, or want to keep billing inside an existing account. The general workflow is:

1. Open the model catalogue in your cloud console and search for Muse Spark Open.
2. Accept the Muse Model License when prompted.
3. Choose a deployment size with enough accelerator memory (see the requirements above) and a region that fits your data-residency needs. For UK and EU organisations, this is the main advantage over a public API.
4. Deploy the endpoint and call it from your application using the provider’s SDK.

Exact screens and instance names differ by provider and change often, so follow your cloud provider’s current documentation for the final steps.

Option 3: Download from Hugging Face and serve with vLLM or SGLang

If you control a suitable GPU server, self-hosting gives you full control over data and costs. The weights are gated on Hugging Face.

1. Sign in to Hugging Face and open meta-models/Muse-Spark-Open-816B-A42B.
2. Complete the access form. It asks for your legal name, date of birth and organisation, and you must be 18 or over.
3. Once approved, log in on your server with a Hugging Face access token.
4. Install a supported inference engine and start the server.

With vLLM, Meta’s model card shows:

pip install vllm
vllm serve "meta-models/Muse-Spark-Open-816B-A42B"

This exposes an OpenAI-compatible endpoint at http://localhost:8000/v1/chat/completions. With SGLang, the equivalent command is:

python3 -m sglang.launch_server \
  --model-path "meta-models/Muse-Spark-Open-816B-A42B" \
  --host 0.0.0.0 --port 30000

The model card also lists Docker Model Runner (docker model run hf.co/meta-models/Muse-Spark-Open-816B-A42B) plus TensorRT-LLM, Transformers, llama.cpp and Unsloth support. Before exposing any self-hosted endpoint to a team, read our guide to AI agent security; an unauthenticated model server on a public port is an easy mistake to make.

Option 4: Run it on Apple Silicon with MLX

Meta’s own demonstration uses two M5 Ultra Mac Studios (512 GB each) running MLX. This is the most compact way to self-host the full model, but it depends on hardware that, at launch, was not yet shipping in that configuration. If you plan to go this way, wait for the 512 GB machines to be available and for community MLX conversions to be tested before buying.

Recommended settings

Meta’s model card recommends temperature 1.0 and top_p 1.0. The model also accepts a reasoning_effort parameter with four levels: low, medium, high and xhigh. The default is xhigh, which gives the most thorough answers but uses the most output tokens.

For everyday tasks such as summarising, drafting or classifying text, try low or medium first. You will usually get faster, cheaper responses. Reserve high or xhigh for multi-step reasoning, coding and agent workflows where accuracy matters more than cost.

Muse Spark Open benchmarks: what Meta reports

Meta published comparisons against other leading open-weight models. These are Meta’s own figures, as reported by VentureBeat; we have not reproduced them, and independent results were not yet available when we checked.

The pattern is mixed rather than a clean win. Muse Spark Open leads on tool use (MCP-Atlas) and the JobBench work-task benchmark, but trails GLM 5.3 and Kimi K3 on the coding-focused Terminal-Bench and DeepSWE tests. If coding is your main use, test it against your own repositories before switching.

A local alternative: Muse Glimmer 30B

If you want a Meta model you can actually run on a desktop or high-end laptop, look at Muse Glimmer 30B, released in August 2026. It is much smaller, has a 131K-token context window and uses the permissive Apache 2.0 licence. The official model card lists approximate memory needs of:

• 64 GB for BF16 weights
• 32 GB for the K-Quant-Dynamic quantisation
• 24 GB for the K-Quant-17GB quantisation

You can serve it with vllm serve "meta-models/Muse-Glimmer-30B" or docker model run hf.co/meta-models/Muse-Glimmer-30B, and quantised versions work in Ollama, LM Studio and llama.cpp. Our guide on how to run an LLM locally walks through those tools step by step.

Using Muse Spark Open with coding agents

Meta says the model works with its own Muse Code agent. Because Muse Spark Open is also available through OpenAI-compatible endpoints, you can point open coding agents that accept custom providers at it, including OpenCode. We compare that approach with Anthropic’s tool in OpenCode vs Claude Code.

A sensible workflow for a small team is:

1. Prototype through OpenRouter with reasoning_effort set to medium.
2. Measure quality and token use on real tasks for a week.
3. Only if volume justifies it, move to a cloud deployment or self-hosting for cost or data-residency reasons.

Limitations to know before you start

• Hardware cost. Self-hosting the full model is out of reach for most individuals and small firms.
• Licence conditions. Revenue thresholds, attribution and use restrictions make it different from Apache-licensed models.
• Text output only. It reads images, video and audio, but it does not generate them.
• Benchmarks are vendor-reported. Treat Meta’s numbers as a starting point, not a verdict.
• Data handling depends on the host. Using a third-party API means your prompts go to that provider. Check its privacy terms and, for UK GDPR, where the data is processed.

FAQ

Is Muse Spark Open free?

The weights are free to download under the Muse Model License, subject to its conditions. Running them is not free: you pay for hardware, cloud compute or per-token API fees.

Can I run Muse Spark Open on my laptop?

No. The full model needs roughly 700 GB of GPU memory even in NVFP4. Use an API, a cloud deployment, or the smaller Muse Glimmer 30B instead.

Where can I download Muse Spark Open?

From the official Hugging Face repository, meta-models/Muse-Spark-Open-816B-A42B, after completing the gated access form. Avoid unofficial re-uploads.

Is Muse Spark Open available in the UK and Europe?

The API partners and cloud catalogues are accessible from the UK and EU, but availability, regions and terms vary by provider. Check the provider’s region list and data-processing terms before sending personal data.

What is the difference between Muse Spark and Muse Spark Open?

Muse Spark is the hosted model used in Meta’s own apps. Muse Spark Open is the downloadable open-weight model that developers can run and build on.

Conclusion

Muse Spark Open gives developers a very large, multimodal, million-token open-weight model with a usable licence for most small organisations. For most readers, the practical path is to try it through OpenRouter or another API partner, compare it with your current model on real tasks, and only consider self-hosting if you have the hardware budget and a clear data-control reason. If you need something you can run on one machine today, start with Muse Glimmer 30B.

Sources: Meta AI: Muse Spark Open announcement; Hugging Face model card and Muse Model License; Muse Glimmer 30B model card; OpenRouter model page; VentureBeat; TechCrunch; The New Stack.