Meta’s Muse Spark Open Model Brings 1M-Token Context Window, Agentic AI Workflows to Apple Mac Studio

Meta on Thursday announced a new open-weight artificial intelligence (AI) model for coding, software development, and complex, multi-step tasks. The model, called Muse Spark Open, is said to have hundreds of billions of parameters and a context window that can handle large amounts of information. Meta has partnered with Apple to bring the model to Mac Studio desktops powered by the M5 Ultra chip. This, as per the company, allows developers to run agentic AI workflows directly on their machines.

Muse Spark Open

In a blog post, Meta described the Muse Spark Open as the most capable US open-weight model to date. The model has 816 billion total parameters, with 42 billion active parameters, and supports a one-million-token context window. It is claimed to be designed for coding and agentic workloads. These include long-running tasks, complex software engineering, and multi-step reasoning.

A notable aspect of the announcement is Meta's collaboration with Apple. Courtesy of this, Muse Spark Open can run locally on two Mac Studio systems powered by Apple's M5 Ultra chip. Meta said that its model has been optimised using Apple's MLX machine learning framework. This, notably, allows developers to run advanced agentic workflows on desktop hardware without depending entirely on cloud-based inference.

As per Meta, the setup takes advantage of Apple's unified memory architecture, which allows the M5 Ultra's integrated GPU to access up to 512GB of memory per Mac Studio. Combined, the setup can accommodate the model's weights using up to 1TB of unified memory. In simple terms, two Mac Studio systems can effectively function as a single supercomputer for running Muse Spark Open.

In addition to Apple's hardware, Muse Spark Open supports several inference engines, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, and llama.cpp. The model is claimed to be capable of running on hardware from AMD, Apple, Intel, and Nvidia. Developers can also customise it using PyTorch and TorchTitan machine learning and deep learning frameworks.

Muse Spark Open introduces custom, auto-tunable Helion GPU kernels in SGLang as well. Meta claims up to 3.95 times kernel-level speed improvements on Nvidia and AMD GPUs. It is also advertised to deliver a 3 percent end-to-end speed improvement on AMD MI350 hardware.

The model is available through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Developers can download its weights through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.

Meta has released Muse Spark Open under its Muse Model License. It permits free usage by developers under certain commercial license terms. On the security front, the company has carried out safety evaluations under its Superintelligence Scaling Framework and thus claims that it meets the thresholds required for release.