What is Muse Spark Open, and can you really run it at home on two Mac Studios?

Yesterday, Meta released the weights of Muse Spark Open, the downloadable version of its Muse Spark line. Weights are the learned numbers that make up a model. Meta ranks it first among open-weight models from the United States. That is Meta's own claim, based on its own benchmark table, and a fuller evaluation report is still to come.

The headline that will catch many readers is that it runs "at home". That is accurate for a particular sort of home, one with two top-end Mac Studios on the desk.

What was released

Muse Spark Open shares its base model with Muse Spark 1.3, according to the PyTorch team, and has had extra safety post-training because an open copy runs without the guardrails Meta applies on its own service.

Meta's model is a mixture of experts, in which only some parameters do the work for each token (a small piece of text). So 42 billion sets the compute per token, and 816 billion sets the memory needed. Note the gate as well: you must fill in a form and be approved, unlike ungated repositories that download straight away.

You can also get it through the Amazon, Google Cloud and Microsoft model catalogues, and hosts such as Together AI, OpenRouter, Groq and Fireworks AI serve it.

What the hardware claim means

Meta's model card asks for at least 700 GB of graphics memory for the NVFP4 version and 2 TB for the full-precision BF16 one. NVFP4 is a compact 4-bit format and BF16 a 16-bit format, so these are two points on the size-and-quality trade-off.

Meta's desktop route uses Apple's unified memory, where the processor and graphics share one pool. Each M5 Ultra Mac Studio can give its graphics up to 512 GB, so two machines hold about a terabyte between them. Meta says the chip's memory bandwidth, 1.2 TB per second, is enough to read the small slice of the model each token needs.

Here's a comparison:

Meta's demonstrations include fixing bugs in an open-source iOS app and driving video-editing software with the model reviewing footage. Those are the company's own examples, so test the responsiveness yourself before you buy anything.

How Meta tuned it for builders

The PyTorch team published details alongside Meta's post. Muse Spark Open ships with kernels, the small routines doing the heavy maths on the chip, written in Helion, a PyTorch language. They report kernel speedups between 1.03 and 3.95 times on NVIDIA and AMD chips, but only a 3 percent improvement overall on AMD's MI350, so the end-to-end gain is much smaller than the top figure suggests.

For tuning, TorchTitan, a PyTorch training platform, adds support for supervised fine-tuning, LoRA adapters (small add-on layers that adjust a model) and reinforcement learning.

Meta's card also says the model is not intended for people under 18, and it recommends human confirmation before irreversible actions when the model operates tools. That advice applies to anyone giving an agent real permissions.

What does the Muse Model License require from you?

The licence is free and starts with broad permission, including fine-tuning and distillation. It is not Apache 2.0, unlike Meta's smaller Muse Glimmer. The conditions that matter:

"Derivatives" is the term to watch. It reaches any AI model built, trained, tuned or distilled with the help of the software or its outputs. So if you use Muse Spark Open's answers to train a small model of your own, that small model comes under the attribution rule too.

The Usage Policy bans things such as creating malware and the unlicensed practice of professions like law or medicine, and the card separately warns against breaking trade-compliance laws. For a Canadian company, none of this requires special steps beyond ordinary legal care, but a firm in a regulated field should read the policy against its actual use. Keep a dated copy of the licence with your project, since your rights rest on the version you accepted.