On Thursday, Meta released a new open-source artificial intelligence (AI) model designed for coding, software development, and complex, multi-step tasks. Dubbed “Muse Spark Open,” the model reportedly features hundreds of billions of parameters and a context window capable of processing vast amounts of information. Meta has partnered with Apple to bring this model to the Mac Studio desktop system, powered by the M5 Ultra chip. According to the company, this enables developers to execute intelligent AI workflows directly on their own machines.
Muse Spark Open
In a blog post, Meta describes Muse Spark Open as the most capable open-weights model to emerge from the U.S. to date. The model boasts a total of 816 billion parameters—with 420 billion active parameters—and supports a context window of one million tokens. It is specifically engineered for coding and “agentic” workloads, including long-running tasks, complex software engineering, and multi-step reasoning.
A key highlight of this release is the collaboration between Meta and Apple. As a result, Muse Spark Open can run locally on Mac Studio systems equipped with Apple’s M5 Ultra chip. Meta noted that the model has been optimized using Apple’s MLX machine learning framework. Notably, this allows developers to run advanced, intelligent workflows on desktop hardware, reducing the need to rely entirely on cloud-based inference.
According to Meta, this approach leverages Apple’s unified memory architecture, allowing the M5 Ultra’s integrated GPU to access up to 512GB of memory on each Mac Studio. Combined, two Mac Studio units can utilize up to 1TB of unified memory to meet the model’s substantial memory requirements. In short, two Mac Studio systems can effectively function as a single supercomputer for running Muse Spark Open.
Beyond Apple hardware, Muse Spark Open supports a variety of inference engines, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, and llama.cpp. The model is reportedly capable of running on hardware from AMD, Apple, Intel, and Nvidia. Developers can also customize it using machine learning and deep learning frameworks such as PyTorch and TorchTitan.
Muse Spark Open also introduces custom, auto-tuning Helion GPU kernels based on SGLang. Meta claims that these kernels deliver speedups of up to 3.95x on Nvidia and AMD GPUs. Furthermore, the company reports a 3% improvement in end-to-end performance when running on AMD MI350 hardware.
The model is available via API platforms such as 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 agreement. This agreement allows developers to use the model free of charge, subject to specific commercial licensing terms. Regarding safety, the company has conducted an assessment based on its “superintelligence scaling framework” and asserts that the model meets all the necessary criteria for release.