Meta has released the weights for Muse Spark Open, which it describes as the most capable US open-weight model available today. The model is notable not just for its scale, but for a specific hardware milestone: it is the first model of its kind that can run entirely on a desktop setup, using two Apple M5 Ultra Mac Studios instead of a data center full of GPUs.

For developers who have grown used to choosing between small local models and massive cloud-hosted ones, this release is an attempt to close that gap. It suggests that frontier-scale, agentic AI could soon be something you run under your desk rather than rent from an API provider.

What Muse Spark Open actually is

Muse Spark Open is a natively multimodal model with 816 billion total parameters, of which 42 billion are active at any given time, and it supports a 1-million-token context window. The “active parameters” figure matters because Muse Spark Open almost certainly uses a mixture-of-experts (MoE) architecture, a common technique where only a fraction of the model’s full parameter count is used to process any given input. This keeps inference costs and memory bandwidth requirements manageable even though the total model is enormous.

Meta says the model was built specifically for coding and agentic work: long-horizon tasks, complex software engineering, and multi-step reasoning, rather than being a general-purpose chat assistant first.

Running a frontier model on two Mac Studios

The headline capability is that Muse Spark Open, optimized with Apple’s MLX framework, can run on two Mac Studios equipped with M5 Ultra chips. This works because of Apple’s unified memory architecture, where the integrated GPU can address up to 512 GB of memory per machine. Combine two machines and you get roughly a terabyte of unified memory, enough to hold the entire model.

The other critical spec is memory bandwidth. Each M5 Ultra offers 1.2 TB/s of memory bandwidth, which Meta says is fast enough to read the slice of the model each token actually needs during generation. The practical result, according to Meta, is that two Mac Studios behave like a single supercomputer dedicated to running this one model.

To demonstrate this, Meta tested agentic workflows running entirely on-device using Muse Code, an assistant powered by Muse Spark Open. Examples included shipping real iOS code changes, bug fixes, and new features to Ice Cubes (an open-source Mastodon client), building the app end to end, and operating DaVinci Resolve video editing software autonomously, reviewing footage frame by frame to identify strong moments and crop them vertically.

How it compares to typical model releases

Most large frontier models today are either closed and accessed only through an API, or open-weight but still realistically require multi-GPU server racks to run at full scale. Muse Spark Open’s pitch is different on two fronts.

First, availability: it is already live through multiple API platforms, including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, and the weights themselves can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. That breadth of distribution, both as a hosted API and as downloadable weights, is broader than many comparable releases.

Second, hardware flexibility: the model runs on a wide set of inference engines, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations across AMD, Apple, Intel, and NVIDIA silicon. Meta also highlights custom, auto-tunable Helion GPU kernels built into SGLang, which it says deliver between 1.03x and 3.95x kernel-level speedups on NVIDIA and AMD GPUs, along with a 3% end-to-end speedup specifically on AMD’s MI350 chip. Developers who want to fine-tune the model for their own use case can do so using PyTorch’s TorchTitan framework.

Licensing, safety, and what it means for developers

The weights are released under what Meta calls the Muse Model License, which it describes as carrying “limited commercial terms” that let many developers use the model for free. The exact boundaries of those commercial terms are not detailed in Meta’s announcement, so developers planning commercial deployments should check the license text directly before building on it.

On safety, Meta says Muse Spark Open received additional safety post-training specifically because open-weight models carry different risks than closed, API-gated ones, since anyone can download and modify the weights once released. Meta states the model was evaluated against its Superintelligence Scaling Framework, covering what it considers the most severe risk categories, and that it met the thresholds required for release. Meta notes a full evaluation report will be available “soon,” meaning independent verification of these safety claims is not yet fully public.

In practice, for a developer today, trying Muse Spark Open means one of two paths: calling it through an existing API provider like Together AI or Groq with no new hardware required, or downloading the open weights from Hugging Face and running them locally if you have access to the necessary hardware, Apple’s M5 Ultra Mac Studio setup being the specific configuration Meta validated for desktop use. For most people without two high-end Mac Studios, the API route remains the practical option, while the local-desktop capability is primarily significant as a proof point of where consumer-grade hardware is headed.

Frequently asked questions

Do I need two Mac Studios to use Muse Spark Open?
No. The model is available through standard API platforms like Groq, Together AI, and OpenRouter, which require no special hardware. The dual Mac Studio setup is specifically for running the full model locally on-device.

What makes a model “open-weight” versus fully open source?
Open-weight means the trained model parameters are publicly downloadable and usable, as Muse Spark Open’s are via Hugging Face and other platforms, but it does not necessarily mean the training data, training code, or full methodology are disclosed.

Why does active parameter count matter if the model has 816 billion total parameters?
Only 42 billion parameters are active per inference, which is typical of a mixture-of-experts design. This keeps computational and memory demands closer to those of a much smaller dense model while still drawing on the knowledge encoded across the full parameter set.

Source: Muse Spark Open