Apple Runs Meta’s Muse Spark Open Across Four Mac Studios
Apple says developers can run Meta’s Muse Spark Open model locally across Mac Studio computers using its MLX framework and RDMA over Thunderbolt. On a cluster…
Apple says developers can now run Muse Spark Open locally on Mac Studio computers, bringing a large open-weight AI model onto desktop hardware rather than requiring cloud data centers. On a cluster of four Mac Studios, the model generates over 36 tokens per second, according to Apple’s published technical results.
The announcement links Apple’s MLX software framework with Meta’s model, offering developers a way to experiment with local inference and agentic applications. Apple describes Muse Spark Open as a frontier-class model with nearly a trillion parameters; its evaluation lists the model at 816 billion parameters.
That is a substantial workload. Apple’s results show how multiple Macs can share the task, while the collaboration highlights a practical trade-off: local operation can reduce reliance on cloud services, but running the model calls for high-end hardware and substantial memory.
Muse Spark Open across connected Macs
Apple’s research post explains how to run the model using the latest version of MLX, which it says was released the same day as the post. The setup starts with a single Mac Studio and scales across several machines using an RDMA over Thunderbolt driver.
Apple’s benchmark covered a cluster of four Mac Studios. The company reported generation throughput above 36 tokens per second, a speed it described as well above average human reading speed.
Apple also compared the M5 Ultra Mac Studio with the previous-generation M3 Ultra. Its research says the M5 Ultra’s GPU Neural Accelerators support matrix multiplication operations used in machine-learning workloads. For time to first token, Apple reported speedups of up to 4.6× against the M3 Ultra.
Generation speed depends more heavily on memory bandwidth, according to Apple. The M5 Ultra delivers 1.2 TB/s of unified memory bandwidth, compared with 819 GB/s for the M3 Ultra. Apple reported 1.6× faster token generation on the M5 Ultra.
The model’s architecture also shapes the computing demand. Apple describes Muse Spark Open as a mixture-of-experts model, which activates only a fraction of its experts during inference. In the same evaluation, the company compared it with Mistral Medium 3.5 and Qwen 3.8, which it identifies as dense models with all parameters active during inference.
Local use comes with a hardware trade-off
CNBC reported that Apple and Meta’s collaboration lets developers run Muse Spark Open on high-end Mac computers for coding, video analysis and other agentic work. The model runs locally rather than in the cloud, which the companies said adds a layer of security and saves users money.
Local inference can keep processing on a user’s own machines, but access is not simply a matter of downloading model weights. CNBC reported that customers can pair between two and four Mac Studio computers and that the model requires Apple’s fastest chips and as much memory as possible.
The cost can be a barrier. CNBC reported that a Mac Studio with the required M5 Ultra chip and 256GB of memory costs at least $10,799. A version with 512GB of RAM was due to be released later that month, and users may need multiple machines for the setup.
Meta AI chief Alexandr Wang told CNBC that many powerful open-weight models are too large to run on most personal computers. He said Apple’s Mac Studios provide enough memory to make Muse Spark Open practical to run.
Developers can download a version of the model from websites such as Hugging Face without extra configuration for Apple silicon, CNBC reported. Meta also tuned Muse Spark Open to run faster on Apple hardware, while Apple published technical documentation on using MLX to split open-source models across multiple machines.
Apple’s benchmark report says Qwen 3.8 and Mistral Medium 3.5 fit on a single Mac Studio with 512GB of memory. For those models, Apple reported memory use of up to 68GB and 158GB, respectively, in its large-context setup.
On a four-machine Mac Studio cluster, Apple reported Muse Spark Open generation throughput of over 36 tokens per second.
Source: machinelearning.apple.com
