What It Actually Takes to Run Muse Spark Open Locally
Meta's 816B Muse Spark Open needs dual Mac Studios or heavy quantization. Here's the real hardware math before you try to self-host it.
What is Muse Spark Open?
Muse Spark Open is Meta’s 816 billion parameter open weights model, released as a mixture of experts architecture with 42 billion parameters active per token. It carries a 1 million token context window and is natively multimodal on the input side, accepting text, images, video, and audio, though it only outputs text. Meta calls it the most capable US open weights model available today, a carefully worded claim, since on the Artificial Analysis Intelligence Index it scores 39, behind several Chinese open models including Mimo 2.6 Pro (46), GLM 5.3 (45), Kimi K3 (44), and Qwen 3.8 (40). Running it locally is less about whether the model is good and more about whether your hardware budget can handle 816 billion parameters.
TL;DR
- Full precision self-hosting realistically means two Mac Studios with M5 Ultra chips and 512GB of unified memory each, a setup Apple reportedly measured at around 25 tokens per second.
- Quantized options exist: Meta shipped an NVFP4 checkpoint at roughly 530GB, and Unsloth released a 2-bit quantization that shrinks the footprint to around 300GB.
- Licensing isn’t Apache 2: Spark Open ships under Meta’s own Muse license, which requires a “built with Muse” attribution and a separate commercial agreement if your model-serving business clears $20 million a year in revenue.
- Cloud hosting is cheaper than buying hardware: multiple providers on OpenRouter, including Together AI, Fireworks, Base Ten, Modal, and Nebius, all serve it at the same $1.25 per million input tokens and $4.25 per million output tokens.
- It’s not Meta’s best model: the closed Muse Wonder and Muse Wonderful models score 52 and 56 on the same index, well above Spark Open’s 39, and the Muse personal agent app runs on a separate customized model entirely.
- Hardware lead times matter: new M5 Ultra Mac Studio hardware capable of running this model was expected to ship with significant delays, pushing availability into early next year for anyone ordering fresh.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How much hardware does Muse Spark Open actually need?
At full precision, an 816 billion parameter model demands an enormous amount of memory just to hold the weights, before accounting for context and overhead. Meta worked directly with Apple to validate a reference setup: two Mac Studios, each configured with an M5 Ultra chip and 512GB of unified memory, networked together to pool enough capacity for the full model. Apple’s own measurement of that configuration came in at approximately 25 tokens per second, a usable but far from blazing speed for a model this size.
That dual-Mac-Studio approach matters because unified memory architectures let Apple Silicon treat RAM and VRAM as a shared pool, which is part of why Mac Studios have become a go-to option for running huge open models without assembling a rack of enterprise GPUs. But two maxed-out M5 Ultra units is not a casual purchase, and new high-end Mac Studio configurations have historically shipped with long lead times, meaning anyone ordering today should expect to wait months before the hardware physically arrives.
What quantized checkpoints are available?
For anyone not buying a pair of Mac Studios, Meta released an NVFP4 quantized checkpoint of Spark Open, which brings the footprint down to roughly 530GB. NVFP4 is a 4-bit floating point format designed to preserve more model quality than naive integer quantization while still cutting memory requirements significantly compared to full precision.
Beyond Meta’s own release, Unsloth produced a 2-bit quantized version that compresses the model further, down to approximately 300GB. That’s a meaningful drop from 530GB, but 2-bit quantization typically comes with some accuracy tradeoff, particularly on tasks requiring precise reasoning or code generation. For most self-hosters, the 2-bit checkpoint is the difference between “needs a serious multi-GPU workstation or high-memory Mac” and “still needs a serious amount of hardware, just slightly less of it.”
Even at 300GB, you’re still looking at a setup that exceeds what a single consumer GPU, or even most prosumer multi-GPU rigs, can hold in VRAM. Running this locally in any quantized form generally means either a high-memory Mac Studio, a multi-GPU server with enough combined VRAM, or CPU offloading setups that will be considerably slower than the 25 tokens per second benchmark Apple measured on its reference hardware.
Is it worth running locally instead of through a hosted API?
For most people building with AI, the answer is no. Muse Spark Open is already available through OpenRouter with multiple backend providers, including Together AI, Fireworks, Base Ten, Modal, and Nebius, all pricing it identically at $1.25 per million input tokens and $4.25 per million output tokens. Notably, Meta itself does not run a hosted version of the open model, only the third-party providers do.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
The economics make the case clearly. Buying two M5 Ultra Mac Studios, each with 512GB of memory, represents a substantial capital outlay before you’ve run a single token. Cloud inference at $1.25/$4.25 per million tokens would take an enormous volume of usage to approach that hardware cost, and that’s before factoring in electricity, maintenance, and the fact that newer, better models will likely arrive before the hardware pays for itself. Local hosting makes more sense for organizations with strict data residency requirements, extremely high sustained usage, or a need to fine-tune and modify the model directly, not for casual experimentation or typical production workloads.
What’s the licensing catch with Muse Spark Open?
Muse Spark Open ships under Meta’s own Muse license rather than a permissive license like Apache 2.0. For most individual developers and small teams, the practical restrictions are minor: you need to display “built with Muse Spark Open” attribution somewhere in your product. But if you’re running a model-as-a-service business generating more than $20 million a year in revenue, you need a separate commercial agreement with Meta before you can serve the model to customers.
This is a meaningful difference from Reflection AI’s competing open release, Beam, a smaller 501 billion parameter model (23 billion active) that ships fully under Apache 2.0. Beam scores lower on the Artificial Analysis Index (30 versus Spark Open’s 39) and trails on several benchmarks, including tool use on MCP Atlas (79 versus 89) and DeepSeek-related evaluations (44 versus 60), though it does edge out Spark Open on terminal bench (80 versus 76). The tradeoff is straightforward: Spark Open is the more capable model but comes with a heavier compute footprint and a more restrictive license, while Beam is smaller, cheaper to run, and cleanly licensed.
How does Muse Spark Open compare to Meta’s other new models?
Spark Open is not Meta’s best model, and it isn’t particularly close. Days before releasing Spark Open, Meta launched two closed frontier models, Muse Wonder and Muse Wonderful, benchmarked against GPT-6 Astra and Claude Opus 5.5. On the Artificial Analysis Index, Wonder scores 52 and Wonderful scores 56, compared to Opus 5.5’s 58. Sitting below those is Muse Spark 1.3 at 48, a closed intermediate model that many expected to be the one Meta would eventually open source. Instead, Spark Open trails that mid-tier closed model by 9 points and trails Meta’s best model by 17.
Access to these closed models is uneven. Muse Spark 1.3 is available on Meta’s API and OpenRouter at $1.25 input/$4.25 output, with a cheaper tier at 10 cents in and 20 cents out if you allow Meta to train on your prompts and outputs. Muse Wonder costs $4 in and $20 out, matching Opus 5.5’s pricing exactly. Muse Wonderful isn’t available through any API at all, it’s locked inside Meta’s Muse Code product on $50 or $200 monthly plans, which limits independent testing of Meta’s strongest claims about the model.
Separately, Meta’s Muse personal agent app, which launched in the US and quickly became a top download on the Apple App Store, does not run on Spark Open. It uses a customized version of the Spark model family, with specifics about which fine-tune or variant kept undisclosed.
Frequently Asked Questions
Can I run Muse Spark Open on a single GPU?
No. Even the 2-bit quantized checkpoint from Unsloth requires approximately 300GB of memory, far beyond what any single consumer or prosumer GPU provides. You’d need a multi-GPU server with substantial combined VRAM or a high-memory unified-memory system like a Mac Studio.
What’s the cheapest way to try Muse Spark Open?
Using a hosted API through OpenRouter is the cheapest practical option. Providers including Together AI, Fireworks, Base Ten, Modal, and Nebius all offer it at $1.25 per million input tokens and $4.25 per million output tokens, with no hardware investment required.
Is Muse Spark Open better than Chinese open models?
Not by the Artificial Analysis Intelligence Index. It scores 39, trailing Mimo 2.6 Pro (46), GLM 5.3 (45), Kimi K3 (44), and Qwen 3.8 (40). Meta’s claim is specifically that it’s the most capable US open weights model, not the best open model overall.
Does Muse Spark Open power Meta’s Muse personal agent app?
No. The Muse app runs on a separate, customized version of the Spark model family, not the publicly released Spark Open weights. Meta hasn’t disclosed exactly which variant or fine-tune powers the agent app.
What license governs Muse Spark Open?
It uses Meta’s own Muse license, not Apache 2.0. Most users just need to include a “built with Muse Spark Open” attribution, but companies running model-as-a-service businesses with over $20 million in annual revenue need a separate commercial agreement with Meta.