Meta has announced the Muse Spark Open model, an open-weight artificial intelligence model with 816 billion parameters that can run entirely locally on two Mac Studio devices equipped with M5 Ultra processors. The model focuses on software development, long-running agentic tasks, and multistep reasoning.
The model uses a mixture-of-experts architecture, activating 42 billion parameters per token, with a context window extending to 1 million tokens. It also supports text, image, video, and audio inputs, enabling its use in tasks that go beyond text generation and conventional coding.
Local Operation Requires Unusual Hardware
Meta developed the model in cooperation with Apple, and it was optimized to run through the MLX machine-learning framework on Mac Studio devices equipped with M5 Ultra processors. The reference configuration uses two devices, each containing 512 gigabytes of unified memory, providing 1 terabyte in total that can accommodate the entire model.
This configuration benefits from the graphics processing unit’s access to the system’s unified memory, as well as the M5 Ultra processor’s memory bandwidth of 1.2 terabytes per second. Meta describes the two devices as operating like a single desktop supercomputer, but these are not typical configurations for users; the Mac Studio version with 512 gigabytes of memory was not available for sale when the article was published, and was expected to launch in late October.
What Does the Model Change for Developers?
Meta tested Muse Spark Open with Muse Code in scenarios based on coding agents. In one demonstration, the model used the Ice Cubes project, an open-source client for Mastodon, to fix bugs and add features. In another demonstration, it controlled DaVinci Resolve, examined video frame by frame, then identified clips suitable for the vertical format and cropped them.
These examples indicate that the model’s focus is not limited to answering questions, but extends to carrying out sequences of actions across multiple tools and programs. The model scored 89.2 points on the MCP-Atlas test and 61.8 points on JobBench, according to the evaluations published by Meta.
Broad Support and a License That Is Not Fully Open Source
Muse Spark Open supports the SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth, and MLX inference frameworks, and can run on AMD, Apple, Intel, and NVIDIA hardware. Its weights are also available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, while Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI provide access through application programming interfaces.
Nevertheless, describing the model as open weight does not mean that it is fully open source. Meta released it under the Muse Model License, which imposes certain commercial restrictions, particularly on high-revenue companies that offer it as a service. At the same time, the license allows free use in a number of commercial scenarios, such as internal use and integration into other products.
Limits of the Superiority Claim
Meta presents the model as the most capable among open-weight models developed in the United States, not among all open models worldwide. The comparisons cited in the article show that GLM 5.3, Kimi K3, and some Qwen models outperform it in certain coding and general-performance tests. Therefore, the practical value of Muse Spark Open is tied to its ability to run locally, its multimodal architecture, and its support for agentic tasks, rather than to a claim of comprehensive superiority over all competitors.