Meta has released the weights for Muse Spark Open, which it calls the most capable open-weight model built in the United States so far. The model has 816 billion parameters, a 1 million token context window, and a rare trick: it is designed to run on Apple silicon desktops. Two Mac Studio machines with the M5 Ultra chip can hold the whole model in memory.
What Muse Spark Open brings
Muse Spark Open is a natively multimodal model with 816 billion total parameters, of which 42 billion are active per token. Meta has aimed it at complex coding tasks and agentic work, the kind of long-horizon jobs where a model has to plan several steps, use tools, and fix its own mistakes. In evaluation results published by Meta, the model claims better scores than other popular open-weight models across agentic, knowledge and reasoning, coding, and tool-assisted multimodal tests.
The model runs in the class of large open-weight releases, and open weights bring real advantages. Developers can run the model on their own hardware, study how it behaves, fine-tune it, and avoid paying per-token API pricing. Meta is shipping the weights under the Muse Model License with limited commercial terms, and many developers can use it for free.
Built for Mac Studio with M5 Ultra
The headline feature is local running on Apple hardware. Two Mac Studios with the M5 Ultra chip, using the latest MLX framework, act like one machine with a terabyte of unified memory. Apple’s unified memory architecture lets each machine’s GPU use up to 512 GB, and the 1.2 TB/s memory bandwidth is fast enough to keep the model fed. You still need serious hardware: at least 1 TB of RAM in total to run it locally.
Meta tested real agentic workflows entirely on-device with Muse Code, its AI coding terminal powered by Muse Spark Open. The model shipped actual iOS changes to Ice Cubes, an open-source Mastodon client, handling bug fixes, new features, and a full build. It also drove DaVinci Resolve end to end, reviewing footage frame by frame to find the strongest moments and cropping them vertically.
Where to get it
Model weights are available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Hosted inference is live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimisations for AMD, Apple, Intel, and NVIDIA silicon. Meta says the open-weight version ships with auto-tunable GPU kernels that deliver up to 3.95x kernel-level speedups on NVIDIA and AMD GPUs.
FAQ
What is Muse Spark Open?
It is Meta’s new open-weight AI model with 816 billion parameters and a 1 million token context window, built for coding and agentic tasks and released with downloadable weights.
Can Muse Spark Open really run on a Mac?
Yes, but you need two Mac Studio machines with the M5 Ultra chip and at least 1 TB of total RAM in all, running through the MLX framework.
Is Muse Spark Open free to use?
The weights are released under the Muse Model License with limited commercial terms, and many developers can use the model at no cost. Hosted API platforms charge their own rates.
How is Muse Spark Open different from Muse Wonder?
Muse Wonder is Meta’s frontier model served in its products. Muse Spark Open is the open-weight release whose weights anyone can download and run on their own hardware.
What can Muse Spark Open do?
It handles multimodal input, long coding sessions, and multi-step agentic work. In Meta’s tests it shipped code changes to an open-source app and edited video in DaVinci Resolve on its own.
