Technology and Scienceupdated 1h ago3 min read
Meta Releases Muse Spark Open With 816B Parameters for Local Mac Studio Execution
This article is an automated aggregation of coverage from news outlets. Our editorials offer in-depth analysis curated by the NewsCord team.
Muse Spark Open has 816 billion total parameters, 42 billion active, and a 1,000,000-token context window. Runs locally on Mac Studio with M5 Ultra, optimized for Macs.
Muse Spark Open Goes Local
Meta released the weights for Muse Spark Open, describing it as its most capable US open-weight model to date, and the model carries 816 billion total parameters with 42 billion active and a 1 million token context window. Meta said Muse Spark Open is natively multimodal with support across text, image, video and audio, and it is built specifically for coding and agentic work, including long-horizon tasks, multi-step reasoning, and complex software engineering. Meta also positioned the release around local execution, saying developers can run the model on a desktop setup built from Apple’s Mac Studio computers, pairing two to four Mac Studio computers with Apple’s M5 Ultra chip.
“Muse Spark Open is available today and brings leading agentic capabilities to your desktop for the first time on Apple M5 Ultra Mac Studios.”
Meta’s announcement tied the on-device pitch to Apple’s MLX framework, and Meta said “together they’re the best way to run a powerful model on your own desktop.” Meta and Apple framed the collaboration as a way to avoid cloud-only use, with Meta releasing the weights and Apple optimizing the Mac Studio path for running Muse Spark Open locally.
Hardware, Benchmarks, and Scaling
Apple’s Apple Machine Learning Research post said MLX on Mac with Apple silicon lets AI researchers and developers experiment locally with the latest models and methods in a private environment. Apple said a cluster of four Mac Studios running Muse Spark Open generates over 36 tokens per second, and Apple described scaling from a single Mac Studio up to multiple Macs using RDMA over Thunderbolt driver. Apple compared Mac Studio generations by evaluating M5 Ultra and M3 Ultra, and Apple reported that TTFT speedups reached “up to 4.6× faster compared to an M3 Ultra.”
Apple also reported generation speed differences by tying them to memory bandwidth, and Apple said the Mac Studio with M5 Ultra reported “1.6× faster tokens/s.” Apple’s post also described memory constraints for Muse Spark Open, saying the model would require more than 1.6TB of RAM in native precision (BF16) for the model parameters alone, while a 4-bit quantized (NVFP4) format lowered the model memory footprint to about 493GB.
Partnership Stakes and Access
Mark Zuckerberg posted on X that “Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple,” and Apple CEO John Ternus replied publicly with a showcase for on-device AI performance. Meta AI chief Alexandr Wang said developers can run Muse Spark Open locally on Apple M5 Ultra Mac Studios, and Wang also described the practical reason that Mac Studios make enough memory available to run powerful open-weight models. Meta said Muse Spark Open is available today through API platforms at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, while Meta said the weights are available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.
“Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.”
Meta said the model is released under the Muse Model License with limited commercial terms, and Meta said the open-weight version ships with custom, auto-tunable Helion GPU kernels in SGLang. Meta also said it evaluated Muse Spark Open in line with its Meta Superintelligence Scaling Framework, and Meta stated that the model met the thresholds the framework requires for release.