Meta's open Muse Spark Open runs on a Mac, if your Mac has about 1TB of memory. Read the license before you build on it

Meta's new open-weight model is real progress, but "runs on a Mac" needs a careful read before you promise it to a client.

What happened

Meta launched Muse Spark Open, which it calls the most capable US open-weight model. It has about 816 billion parameters, around 42 billion active per token, and a context window up to one million tokens. It accepts text, images, video and audio, and Meta built it for coding and agentic tasks. The weights are on Hugging Face, and TechCrunch reports it is available through several API platforms.

Meta and Apple worked on running it locally on Macs. The Decoder reports it runs on two Mac Studios with 512GB of unified memory each. TechCrunch, citing Engadget, says local use needs at least 1TB of RAM. Hugging Face tested a 2-bit quantized version on a 512GB Mac M3 Ultra at 18 tokens per second. Hugging Face also notes reasoning cannot be turned off; the lowest setting is "low".

On Meta's benchmarks it leads open rivals like Qwen 3.8, GLM 5.3 and Kimi K3 in several categories, falls behind some on coding, and stays well short of Meta's paid Muse Wonder Full. The Decoder reports the license is free to use but requires the label "Built with Muse Spark Open", and providers earning over $20 million a year need a separate agreement to offer it as an API.

My take

Open weights give you two things: you can run the model where the data lives, and you are not tied to one vendor's pricing. Both matter for client work.

But be honest about the hardware. Roughly 1TB of memory is not a laptop. For most small teams, the realistic path is calling it through an API provider, which is still useful because you can switch providers later.

Before you put it in a client workflow:

• Read the license, including the labeling rule, and check how it applies to what the client ships.
• Budget for reasoning tokens, since you cannot switch reasoning off.
• Run your own test set. Benchmark wins in "agentic" categories do not tell you how it handles your CRM data.
• Keep the model behind one interface in your automation so swapping it later is a config change.