Meta Muse Spark Open: Specs, Hardware Requirements
Open-weight models usually force a trade-off: either they’re small enough to run at home, or big enough to be impressive. Meta’s new Muse Spark Open tries to sit in the awkward middle. It’s huge on paper, but it only uses a small slice of itself for each token, and that changes what hardware it needs.
The short version: Muse Spark Open is an open-weight multimodal model with about 816 billion total parameters and about 42 billion active per token. It accepts text, images, audio and video, outputs text, and handles up to 1,048,576 tokens of context. Meta built it mainly for coding, tool use and agent-style tasks.
Muse Spark Open at a Glance
| Spec | Details |
|---|---|
| Total parameters | ~816 billion |
| Active parameters per token | ~42 billion |
| Architecture | Mixture of Experts (MoE) |
| Experts per MoE layer | 256 |
| Experts used per token | 8 |
| Input | Text, images, audio, video |
| Output | Text |
| Context window | Up to 1,048,576 tokens |
| Checkpoints | BF16 and NVFP4 |
What “816B Total, 42B Active” Means
In a Mixture-of-Experts model, each layer has many small specialist networks. Here there are 256 of them per layer, and for every token the model picks just eight.
So you get the knowledge capacity of an 816B model, while each token only costs about as much compute as a 42B one. That’s roughly 5% of the parameters doing the work at any moment.
There’s a catch worth knowing: all the weights still have to sit in memory, even the ones that aren’t used for a given token. Fast to run doesn’t mean small to store. That’s why the hardware numbers below look the way they do.
A Context Window Shared Across Everything
The 1M-token window isn’t just for text. It’s shared between text, images, audio and video. The model has separate encoders for visual and audio input, and its audio system handles 24kHz sound at about 12.5 tokens per second.
In practical terms, that means a long audio clip or video eats into the same budget as your written prompt. If you plan to feed it hours of media, do the token math first.
It also supports tool calling and several reasoning levels, from low to high, so you can trade speed against depth.
What Meta Showed It Doing
Meta’s demos all lean toward agent-style work:
- 3D coding: The model built and improved Three.js objects, using browser screenshots to judge its own output and then editing the code.
- Browser control: It operated a browser from screenshots alone, without being given the page’s HTML or DOM.
- Voice shopping: A user described a product out loud, and the model combined audio, vision and agent tools to find a match.
These are demonstrations, not independent benchmarks, so treat them as a sign of intent rather than proof of reliability.
Can You Run Muse Spark Open Locally?
Yes, but not on an ordinary laptop or gaming PC.
On Apple Silicon, the model runs through llama.cpp, oMLX and mlx-lm on high-end Macs:
| Quantization | Memory used | Speed (M3 Ultra, 512GB) |
|---|---|---|
| 2-bit | ~304GB | ~18 tokens/sec |
| 3-bit | ~344GB | ~21 tokens/sec |
Even heavily compressed, it needs more than 256GB of memory.
On Nvidia GPUs, the BF16 version runs across multiple H200 systems, while the NVFP4 version supports newer Blackwell chips like the B200 and GB300. Supported frameworks include SGLang, vLLM and Hugging Face Inference Endpoints.
Who Should Care About This Model?
- Developers building agents who want an open-weight model with tool use and vision built in.
- Teams with strict data rules who want to host a model themselves instead of sending data to an API.
- Researchers who need long-context, multimodal experiments without being locked to one provider.
