Muse Spark Open: What It Takes to Run Meta’s 816B Open Model, and What the License Allows
Muse Spark Open is Meta’s new open-weight AI model, released on October 15, 2026. It has 816 billion parameters, a 1-million-token context window, and Meta says it can run entirely on two Apple Mac Studios with the M5 Ultra chip. Meta calls it “the most capable US open-weight model to date.”

This guide covers what our video promised: the hardware you need, and the license terms you should read before you build anything with it.
Key Takeaways
- Release date: October 15, 2026, from Meta Superintelligence Labs.
- Size: 816B total parameters, about 42B active per token, 1M-token context window.
- Inputs: text, image, video and audio (natively multimodal).
- Local hardware: two M5 Ultra Mac Studios with 512GB each, or GPUs with at least 700GB of VRAM for the NVFP4 version (2TB for BF16).
- Price: the weights are a free download on Hugging Face; hosted APIs are available from Groq, OpenRouter, Together AI and others.
- License: the Muse Model License is free for most uses, but large “model as a service” businesses must sign a separate deal with Meta.
Watch: our 30-second video summary (in Vietnamese).
What Is Muse Spark Open?
It is the open-weight version of Meta’s Muse Spark model. “Open-weight” means you can download the trained model files and run them on your own machines, instead of only reaching the model through someone else’s server.
It is a sparse mixture-of-experts (MoE) model. An MoE model splits its knowledge across many smaller “expert” networks and only switches on a few of them for each token. That is why the model has 816 billion parameters in total but uses only about 42 billion for each token it generates. The result is a very large model that is cheaper to run than its size suggests.
Meta built it for coding and agentic work. An “agent” is a model that does more than chat: it plans steps, calls tools, checks the result and fixes its own mistakes. According to Meta, it handles long-horizon tasks, complex software engineering and multi-step reasoning.
Meta added extra safety post-training because the open version can run without Meta’s usual guardrails.
How Big Is It? Specs and Real-World Tests
The sources we read do not include benchmark scores. Meta points to a separate evaluation methodology report and model card for those numbers. Here are the published specs instead:
| Spec | Value |
|---|---|
| Total parameters | 816B |
| Active parameters per token | ~42B |
| Context window | 1M tokens |
| Input types | Text, image, video, audio |
| Number of experts | 256 |
| Released checkpoints | BF16 and NVFP4 |
| Developer | Meta Superintelligence Labs |
What Meta tested on-device
Meta ran several agent workflows entirely on two Mac Studios, using its Muse Code tool powered by the model:
- Shipping iOS code: the model made real changes to Ice Cubes, an open-source Mastodon app, including bug fixes, new features and a full build.
- Editing video: the model drove DaVinci Resolve end to end, reviewing footage frame by frame to find the strongest moments and crop them into vertical clips.
These are Meta’s own demos, not independent tests.
Muse Spark Open Requirements: The Hardware You Need
This is the part most people will hit first. The model is big, and “runs on a desktop” still means a very expensive desktop.
| Setup | What you need | Notes |
|---|---|---|
| Two Mac Studios (MLX) | 2 × Mac Studio with M5 Ultra and 512GB unified memory | Runs the NVFP4 checkpoint; Meta’s recommended desktop setup |
| Four Mac Studios (MLX) | 4 × Mac Studio with 256GB unified memory (1TB total) | Same NVFP4 checkpoint, split four ways |
| GPU server, NVFP4 | At least 700GB of VRAM | For example a four-GPU NVIDIA Blackwell node with TensorRT-LLM |
| GPU server, BF16 | At least 2TB of VRAM | Full-precision weights |
| Single workstation (Unsloth) | About 300GB of combined RAM and VRAM, or unified memory | Smallest 2-bit Dynamic GGUF; lower precision trades away some quality |
Why two Mac Studios work
Apple’s unified memory lets the M5 Ultra’s built-in GPU use up to 512GB per Mac Studio. Link two machines with Thunderbolt 5 and RDMA, and MLX splits every layer of the model across both. Each Mac holds half the weights and does half the math for every token. Only small intermediate results travel between the machines. Meta says the pair “behave like one supercomputer with a terabyte of unified memory,” and the M5 Ultra’s 1.2 TB/s of memory bandwidth is fast enough to read the small slice of weights each token needs.
Supported software
The model had day-one support in SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA chips. It also ships with two “draft models” (DFlash2 and DSpark) for speculative decoding, a trick where a small model guesses several tokens ahead and the big model checks them in one pass, which speeds up output.
Muse Spark Open License: What You Can and Cannot Do
The weights are released under the Muse Model License. Meta describes it as having “limited commercial terms, which allow many developers to use the model for free.” Here is what the license text on Hugging Face says in practice:
| Question | Answer under the Muse Model License |
|---|---|
| Can I download and use it for free? | Yes. Permission is granted free of charge to use, copy, modify, fine-tune, distill and distribute it. |
| Can I use it in a commercial product? | Yes, including as a feature inside your own product. |
| Do I need to credit Meta? | Yes. You must prominently show “Built with Muse Spark Open” in the product’s interface, documentation, website or settings menu. |
| Is there a revenue limit? | Only for “model as a service” businesses (selling access to AI models via API) with more than $20 million in gross revenue over any 12 months. They must notify Meta and sign a separate agreement. |
| Who owns the outputs? | Meta claims no ownership of outputs. |
| Age limit? | The model is not intended for people under 18. |
The catch: the attribution rule also covers “Derivatives,” which includes any model trained, fine-tuned or distilled using Muse Spark Open or its outputs. If you use it to generate training data for another model, that model falls under the same terms. You also need to log in and accept the conditions on Hugging Face before downloading.
How Do You Get Access?
You have three main options:
- Use a hosted API (no hardware needed). The model is available from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. This is the easiest way to try the model before buying anything.
- Use a cloud model catalog. Teams that want a private instance can deploy it from Amazon SageMaker AI, Google Cloud Model Garden or Microsoft Foundry Models.
- Download the weights. Get the BF16 or NVFP4 checkpoints from Hugging Face, then run them with MLX on Mac or with SGLang, vLLM, TensorRT-LLM, llama.cpp or Unsloth on PCs and servers. Unsloth Desktop runs on macOS, Windows and Linux and can serve the model through OpenAI- and Anthropic-compatible endpoints for your coding agents.
For fine-tuning, Meta supports PyTorch’s TorchTitan (including LoRA and reinforcement learning), NVIDIA NeMo AutoModel, Hugging Face’s SFTTrainer and RadixArk’s Miles framework.
To try local AI on a smaller scale first, see our n8n and Ollama workflow and Ollama prompt templates.
Is Meta’s Open Model Worth It?
Pros
- One of the largest models you can legally download and run yourself, with a 1M-token context.
- Built for agentic coding, with real on-device demos from Meta.
- Broad launch support: almost every major inference engine and chip vendor.
- Free for most developers, and outputs belong to you.
Cons
- Hardware is out of reach for most individuals: even the cheapest local route needs about 300GB of memory, and the recommended Mac setup is two top-end Mac Studios.
- No benchmark numbers in the launch posts we read; you have to dig into Meta’s separate report.
- Mandatory “Built with Muse Spark Open” branding, plus extra terms for large API businesses.
For most people, the practical move is to test it through a hosted API. Buy hardware only if you have a clear reason to keep data on your own machines. Developers comparing coding agents may also want our Claude Code vs Codex comparison.
Muse Spark Open FAQ
What is Muse Spark Open?
Muse Spark Open is Meta’s open-weight AI model with 816B total parameters (42B active), a 1M-token context window, and text, image, video and audio input. It was released on October 15, 2026.
Can I run Muse Spark Open on one Mac?
Not at full NVFP4 quality. Meta’s setup uses two M5 Ultra Mac Studios with 512GB each. A single machine with about 300GB of combined memory can run Unsloth’s 2-bit version.
Is Muse Spark Open free?
Yes. The weights are free to download and use under the Muse Model License. You must display “Built with Muse Spark Open”, and API businesses earning over $20 million a year need a separate agreement with Meta.
Where can I use Muse Spark Open without buying hardware?
Through hosted APIs such as Groq, OpenRouter and Together AI, or the AWS, Google Cloud and Microsoft model catalogs.
What is Muse Spark Open good at?
According to Meta, it is built for coding and agentic tasks: long-horizon work, complex software engineering and multi-step reasoning.
Sources
- Meta Research: Muse Spark Open
- Meta for Developers: Build with Muse Spark Open
- Hugging Face: Muse-Spark-Open-816B-A42B model card and license
- 9to5Mac: Meta launches new Muse Spark Open model optimized to run on Mac Studio
Last updated: October 16, 2026