Thought Leadership
The model with 816 billion parameters is optimized for local operation on two Mac Studios with 512 GB of memory each.
Meta has released the AI model Muse Spark Open. The model weights are freely available. Meta describes it as the most capable open-weight model from the United States to date. Just a few days earlier, the company introduced Muse Wonder, a closed, paid frontier model.
Muse Spark Open has 816 billion parameters, 42 billion of which are active per token. The context window spans one million tokens. The model is tailored for programming and agentic tasks. According to Meta, its core strengths lie in long-horizon tasks, multi-step reasoning, and software development. Independent benchmarks are not yet available.
Two Mac Studios as One System
Meta has optimized the model for local operation on two Mac Studios equipped with Apple’s M5 Ultra chip and the MLX framework. Thanks to the unified memory architecture, the integrated GPU of the M5 Ultra can access up to 512 GB of RAM per machine, according to Meta. Two devices can thus accommodate the entire model. The chip’s memory bandwidth of 1.2 TB/s is sufficient to read the required portion of the model for every token.
“The result is that two Mac Studios behave like a supercomputer with one terabyte of unified memory.”
Meta
Mark Zuckerberg described the setup as the best way to run a high-performance model right on one’s desk.
Mac Studio with 512 GB Not Yet Available
Apple has not yet launched the Mac Studio with 512 GB of memory. According to MacRumors, it is expected to be released later in October. In the US, the 256 GB configuration already costs more than $10,000.
The number of highly capable open-weight models has grown recently. On October 5, US startup Reflection AI unveiled its model Beam, followed on October 6 by French AI firm Mistral with its one-trillion-parameter model Large 4. According to the US National Institute of Standards and Technology (NIST), the Chinese model GLM-5.3 by Z.ai is currently regarded as the most capable open-weight model for cyber tasks.
(Editorial Team)
