Running Meta's Muse Spark Open Locally: What It Actually Takes
Meta's open-weight Muse Spark Open is an 816B parameter model built for Mac Studio. Here's the hardware, memory, and setup reality.
What is Muse Spark Open?
Muse Spark Open is Meta’s open-weight large language model, released as a downloadable alternative after the company’s new flagship, Muse Wonder, took over as its best internal model. Meta describes it as the most capable open-weight model out of the US, a careful phrase that concedes some Chinese open-weight models may still beat it on raw capability. The model weighs in at 816 billion total parameters with 42 billion active at a time (listed on Hugging Face as “816B-A42B”), meaning it uses a mixture-of-experts architecture rather than running all parameters on every token. It’s tagged on Hugging Face as an image-text-to-text model, so it handles multimodal input, not just plain text chat.
TL;DR
- Muse Spark Open is Meta’s downloadable open-weight model, demoted from flagship status once Muse Wonder took over as the company’s top internal model.
- The model uses a mixture-of-experts design with 816 billion total parameters and 42 billion active per token, shipped across 65 safetensors shards on Hugging Face.
- Meta explicitly optimized the model to run on Apple Mac Studio hardware, specifically machines built around the M5 Ultra chip.
- Even compressed FP4 variants reportedly need around 700GB of unified memory or VRAM, with full precision pushing closer to 2TB, putting this far outside consumer GPU territory.
- Access is gated on Hugging Face (“manual” approval), not an instant download, and the license carries Meta’s usual custom terms and use policy rather than a fully permissive open license.
- In practice, almost nobody will run this on a single machine at home. The realistic audience is labs, researchers, and firms with serious multi-GPU or high-memory Apple hardware, plus people building distilled or quantized derivatives.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Why did Meta release an open-weight version at all?
Meta’s pattern with its Muse line has been to ship a closed, state-of-the-art model first, then release an open-weight sibling once a newer flagship takes its place. Muse Wonder arrived as Meta’s new top-tier model, scoring competitively with Claude Opus 5.5 and GPT 6.1 class models on benchmarks like the Super Intelligence Index and coding evaluations, while Muse Spark, the model it replaced at the top, got pushed out publicly as Muse Spark Open. That’s the same playbook plenty of labs use: once the frontier model ships, the previous generation becomes the goodwill release.
The framing matters here. Meta calls it the most capable open-weight model “to date” from a US lab, which is a narrower claim than “the best open-weight model, period.” Open-weight releases out of Chinese labs, including various Qwen and DeepSeek-class models, have been competitive or ahead on specific benchmarks for a while. Meta’s claim is really about leadership among US-based open releases, not a universal crown.
What are the actual hardware requirements?
This is where most people’s interest in “running it locally” runs into reality. At 816 billion total parameters, even with a mixture-of-experts structure that only activates 42 billion parameters per forward pass, the full weight set still has to live in memory somewhere, because the router can call on any expert at any time.
Reported figures put even an aggressively quantized FP4 version at around 700GB of memory, and higher-precision variants closer to 2TB. That immediately rules out anything resembling a normal gaming PC or workstation GPU. A single high-end consumer card tops out nowhere close to that, and even multi-GPU servers need serious aggregate VRAM to hold it.
This is also why Meta’s collaboration with Apple matters. Apple Silicon’s unified memory architecture lets the CPU and GPU share one large memory pool instead of being capped by a discrete GPU’s onboard VRAM. A Mac Studio built around the M5 Ultra chip can be configured with very large amounts of unified memory, which makes it one of the few plausible single-machine options for holding a model this size without resorting to a server rack of GPUs. That’s the specific hardware angle Meta optimized for: not “any Mac,” but the top-end Mac Studio configuration.
Is it realistic to run Muse Spark Open at home?
For nearly everyone, no. Even the most generously configured Mac Studio sits well below the 700GB-plus memory footprint the quantized model reportedly needs, and getting there at all means buying the most expensive Apple workstation configuration available, not a typical consumer purchase. For comparison, the kind of enterprise or research hardware this model targets, multi-GPU clusters or the highest-memory Mac Studio tier, is priced and positioned for labs and companies, not hobbyists.
What’s realistic instead:
- Distillation. Smaller models trained to mimic Muse Spark Open’s behavior are already appearing from the community, trading some capability for something that fits on a single GPU or a modest Mac.
- Quantized community builds. Expect further compressed versions (below FP4) to show up on Hugging Face as people try to shrink the memory footprint further, likely with real quality trade-offs.
- Cloud or rented hardware. Firms that want to run the full model without buying dedicated infrastructure can rent high-memory instances instead of owning hardware outright.
- API access to Meta’s hosted models. If the goal is just using Meta’s best capability rather than running weights yourself, Meta’s hosted Muse Wonder (the closed flagship) is the more practical route.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
How do you actually get the weights?
Muse Spark Open is listed on Hugging Face under the meta-models organization, split across 65 safetensors shards with a standard model.safetensors.index.json manifest. Access is gated with manual approval rather than an open download, so getting the files requires requesting access and agreeing to Meta’s license and use policy, which are Meta’s own terms rather than a fully permissive open-source license like Apache or MIT. That distinction matters for anyone planning to fine-tune or redistribute derivatives commercially: read the license and use policy before assuming open weights means unrestricted use.
Once access is granted, loading the model requires the transformers library (it’s tagged as compatible) along with infrastructure capable of sharding the model across enough memory, whether that’s a multi-GPU setup, a cloud instance with sufficient aggregate VRAM, or a maxed-out Mac Studio.
How does it compare to Meta’s own flagship and competitors?
Muse Spark Open is explicitly the step below Meta’s current best model, Muse Wonder, which itself scored well against Claude Opus 5.5, GPT 6.1 class models, and Google’s Gemini 4 on several benchmarks, including a strong coding score on SWE-bench style evaluations. Muse Spark Open doesn’t carry that same frontier status. It’s the previous generation, now handed to the open community.
Against other open-weight efforts, the honest comparison is messier. Alibaba’s Qwen line and other Chinese labs have shipped competitive open models for specialized tasks like image generation and fast inference, often explicitly optimized to run on much more modest consumer hardware (some image models in this space target cards as ordinary as an RTX 4070). Muse Spark Open takes the opposite path: it’s a large, dense-capability language and multimodal model that favors capability over accessibility, which is consistent with Meta’s “most capable” framing rather than a “most runnable” one.
Frequently Asked Questions
How many parameters does Muse Spark Open have?
816 billion total parameters, with 42 billion active per token due to its mixture-of-experts architecture, according to its Hugging Face model card (listed as 816B-A42B).
Can I run Muse Spark Open on a regular gaming PC?
No. Reported memory requirements run from roughly 700GB for a compressed FP4 version up to around 2TB for higher precision, far beyond what any consumer GPU or standard workstation offers.
Why did Meta optimize it for Mac Studio specifically?
Apple Silicon’s unified memory architecture lets a single machine share a large memory pool between CPU and GPU, which is one of the few ways to approach this model’s footprint without a multi-GPU server setup. Meta worked with Apple to optimize the model for Mac Studios built around the M5 Ultra chip.
Is Muse Spark Open fully open source?
The weights are downloadable from Hugging Face, but access is gated behind manual approval, and usage is governed by Meta’s own license and use policy rather than a permissive open-source license like Apache 2.0.
How does Muse Spark Open compare to Meta’s Muse Wonder model?
Muse Spark Open is the previous generation. Muse Wonder is Meta’s current flagship and benchmarks competitively with Claude Opus 5.5 and other frontier closed models, while Muse Spark Open was released openly after being superseded internally.
