Muse Spark Open is an open-weight multimodal model with 816B total parameters and 42B active parameters, along with a 1M-token context window. It is designed for coding, agentic work, and multi-step reasoning.
Its standout feature is that it can run on a desktop using two Apple M5 Ultra Mac Studios through MLX. It also supports inference engines such as SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, and Unsloth. The model can also be further fine-tuned with PyTorch.
Two Mac Studios Combined into One System
Each Mac Studio powered by the M5 Ultra chip provides up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth. When two machines are used together, the system can accommodate the full model and operate like a supercomputer with terabytes of unified memory.
When a Model Is Too Large to Run on a Typical Computer
Running advanced models on a desktop is not easy, especially Muse Spark Open, which has 816B total parameters and supports a 1M-token context window.
Muse Spark Open addresses this challenge by running on two M5 Ultra Mac Studios, allowing users to run the model locally for coding, complex software projects, and multi-step agentic tasks.
Where Muse Spark Open Fits Among Meta’s Models
Meta launched Muse Spark Open as an open-weight model that the company described as the most capable U.S.-based model at the time. Its main goal is to give users access to an advanced model that can run in their own environments.
The model is designed for coding, agentic work, and tasks requiring multi-step reasoning. It supports long-running tasks and complex software engineering. This concept aligns with Meta’s goal of expanding access to personal superintelligence.
How Were Previous Models Discussed in This Launch?
The source does not specify the names, versions, sizes, or test results of previous models, so comparisons can only be made against the information announced for Muse Spark Open.
| Factor | Previous model | Muse Spark Open |
|---|---|---|
| Model details | Not specified in the source | Launched as an open-weight model |
| Parameters | Not specified in the source | 816B total and 42B active |
| Context window | Not specified in the source | 1M-token |
| Desktop operation | Not specified in the source | Runs on two Apple M5 Ultra Mac Studios |
From Coding to Video Editing: What Can It Actually Do?
Muse Spark Open is suitable for coding, complex software engineering, and long-running agentic tasks that require multi-step reasoning.
In Ice Cubes, the model can fix bugs, add features, and complete the iOS app build process from start to finish.
For video work, the model drives DaVinci Resolve end to end, inspecting footage frame by frame to find the best moments and then cropping them into a vertical format.
Developers who want to adapt the model to specialized tasks can use PyTorch and TorchTitan for additional tuning.
Other Options for Running and Using the Model
Muse Spark Open offers several usage options depending on the developer’s environment.
| Factor | Muse Spark Open on Mac Studio | SGLang or vLLM | API providers |
|---|---|---|---|
| Access | Runs on a desktop through Apple MLX | Runs through an inference engine | Used through an API |
| Supported hardware | Apple silicon | AMD, NVIDIA, and other hardware supported by the engine | No need to manage the serving hardware yourself |
| Usage channel | Download weights from Hugging Face | Use the weights with SGLang or vLLM | Together AI, Fireworks AI, Baseten, Modal, Groq, Nebius, and OpenRouter |
If you have a Mac Studio and want to run the model locally, MLX is the most direct option. Those who already have an inference system can use SGLang or vLLM, while APIs are suitable for getting started without managing model serving yourself.
Strengths Highlighted at Launch and Limitations to Watch
Muse Spark Open stands out with 816B total parameters, 42B active parameters, and a 1M-token context window. It is suitable for coding, software engineering, and multi-step tasks. It supports SGLang, vLLM, MLX, llama.cpp, and other engines across hardware from multiple vendors. It can also be further customized with PyTorch.
The limitation is that, according to the launch information, desktop operation requires two Mac Studios with M5 Ultra chips. The Muse Model License also includes limited commercial terms, and open-weight models still require careful attention to safety, even after safety post-training and model-release criteria have been completed.
Pros
- +1M-token context window
- +Supports multiple types of hardware and inference engines
- +Can be further customized with PyTorch
- +Runs on desktops through Mac Studio and MLX
Cons
- −Desktop operation requires two Mac Studios
- −The Muse Model License has limited commercial terms
- −Open-weight models still involve safety considerations that require monitoring
Costs Go Beyond Downloading the Model Weights
Running Muse Spark Open on a desktop requires two M5 Ultra Mac Studios. The burden therefore extends beyond downloading the model weights to acquiring and maintaining the machines.
If you do not want to manage the system yourself, you can use APIs from providers such as Baseten, Fireworks AI, Groq, or Together AI. However, this creates dependence on external services, and the source does not specify pricing.
For self-hosting, you still need to install and maintain an inference engine such as SGLang, vLLM, or MLX that is suitable for your hardware. Commercial use also requires carefully reviewing the Muse Model License because it includes limited commercial terms.
Which Teams Is It Suitable For, and Who Should Wait for More Information?
Made for
- Developers and teams working on coding or long-running agentic tasks
- Teams that want to customize the model with PyTorch and have infrastructure for running large models
Think twice
- Teams that need to review additional evaluation and safety details
Skip this one
- People without two Mac Studios — they should wait for information about use on other hardware
Those who want to make decisions based on complete evaluation data should wait for a more detailed model card and evaluation report before using it in production.
Muse Spark Open should be viewed as an ecosystem, not merely a parameter count, because the user experience depends on the model, hardware, inference engine, and license.
Its value will become clearer as people customize the model and use it to create real-world work in different forms. Continue following real-world usage, and if you create something with Muse Spark Open, remember to share it through the Hugging Face Forums — the team is already welcoming contributions from the community.