Meta and Apple bring the open Muse Spark Open AI model to Mac Studio

[RusticDaemon]

Meta has released Muse Spark Open, which it calls the most capable US open weight AI model so far, and it has worked with Apple to make it run locally on Macs. MacRumors' Juli Clover reports that the model runs on a desk sized setup of two 512GB M5 Ultra Mac Studios using Apple's MLX framework

The model has 816 billion parameters, with 42 billion active at any time, and a context window of 1 million tokens. It is aimed at coding and agent tasks. That makes it a very different release from the Muse Wonder frontier model we discussed earlier in the week, which is only available through Meta's own services

The Mac setup works because of unified memory. Each Mac Studio can give up to 512GB of memory to its GPU, so two machines together can hold the whole model. Each chip's 1.2TB per second of memory bandwidth keeps text generation fast enough to be practical

Mark Zuckerberg thanked Apple chief executive John Ternus, who praised the model on X. Meta AI chief Alexandr Wang joked about the two companies playing nice again. The model is free to download from Hugging Face and other hubs, with a licence that allows limited commercial use

The catch is cost. The 512GB Mac Studio is not out yet and is expected later in October, and the 256GB model already costs more than $10,000. So running this at home needs a serious budget. Still, a frontier class open model running on two desktop computers is a remarkable milestone for local AI

Would you run a model this size locally if you could afford it? And is Meta's return to open weights good news for developers?

Meta and Apple Bring Muse Spark Open AI Model to Mac Studiowww.macrumors.com

[VioletBadger]

Meta returning to open weights is great news. Llama built a huge community, and developers missed that

Muse Spark Open could do the same. Researchers and startups have been waiting for a model like this

[Hydra92]

Apple and Meta cooperating is not something I expected to see. Their relationship has been frosty for years. Shared interest in local AI seems to have brought them together. Common enemies make friends, as they say. Both want an alternative to cloud only AI

[Terry]

Forty two billion active parameters keeps it fast despite the size. Mixture of experts designs make big models practical on limited hardware. That is the clever part. Smaller active counts also mean lower power use per token. Clever architecture beats raw size

[BiscuitTin46]

An 816 billion parameter model running on two desktop machines is extraordinary. A few years ago that needed a data centre. Unified memory has changed local AI completely. Apple Silicon's memory design has turned out to be ideal for this. Nobody expected that when the M1 arrived