Meta put out Muse Spark Open today and called it the “most capable US open-weight model to date.” The second claim is the one worth chewing on: Meta says it’s the first model in its class that can run on your desktop, specifically two Apple M5 Ultra Mac Studios using the latest MLX framework. Engadget pinned the requirement at 1TB of RAM. So, two machines, both configured well past anything a normal person buys, wired together to serve one model.
Call that local inference if you want. It is technically true and practically meaningless for almost everyone reading about it.
The partnership matters more than the hardware flex
What’s actually new here is Apple showing up as a named partner on a frontier-adjacent open-weight release. MLX has been a quietly excellent piece of software for a couple of years, mostly carried by hobbyists quantizing other people’s weights and posting throughput numbers. Having Meta ship a model with MLX as a first-class target, rather than something the community retrofits three weeks later, changes the status of Apple silicon as an inference platform. It stops being a clever hack and starts being a supported path.
Meta gets something out of this too. Every open-weight release in the past two years has implicitly been a story about Nvidia: how many H100s you need, how much VRAM, which cloud will rent it to you. A launch where the hero configuration is Apple hardware sidesteps that framing entirely. Unified memory is the one thing Apple has that datacenter GPUs don’t, and a terabyte of it on a desk is genuinely hard to replicate at any sane price elsewhere.
”Open-weight” is doing a lot of work
The label has drifted. A model you can download but cannot run is open in roughly the way a blueprint for a bridge is open. Useful to researchers, to people building quantizations, to anyone with a rented cluster. Not useful to the developer who wanted to stop paying per token.
That said, the first useful thing about these releases is rarely the flagship weights. It’s what the community does to them in the following month. A 1TB model gets squeezed, pruned, distilled and quantized by people who have one Mac Studio, or a 128GB MacBook Pro, and a lot of patience. If Meta’s MLX work holds up, that squeezing happens on a much better foundation than usual.
The missing number is performance. Nobody has published tokens per second on the two-Studio setup, and that figure decides whether this configuration is a real deployment target or a press-release diorama. A model that answers at reading speed on hardware you already own is interesting. One that takes a minute per response is a benchmark entry. Meta hasn’t said which it is.