Meta and Apple have collaborated to make Muse Spark Open run on Mac, using two Mac Studio computers with M5 Ultra chips and Apple Silicon’s MLX framework. The key highlight is that it is a multimodal model with a total of 816 billion parameters, supporting long-horizon tasks, complex software development, and multi-step reasoning.
However, the source does not yet provide details about pricing, real-world performance, or comparisons with competing models, so it is still impossible to determine how worthwhile it is for general users.
Opening image of Muse Spark Open on Mac Studio
This image depicts Muse Spark Open running on a desktop using two Mac Studio computers with M5 Ultra chips, along with the MLX framework on Apple Silicon. The key point is that Meta says the model can run locally, while Apple supports running large models on Mac Studio.
When large AI models move onto the desktop
Bringing a large AI model to run on a desktop is no small matter, especially when Muse Spark Open has a total of 816 billion parameters and must handle complex, multi-step tasks.
What is particularly interesting is that Meta says Muse Spark Open can run on two Mac Studio computers equipped with M5 Ultra chips through Apple’s MLX framework. This collaboration brings models of this scale closer to the desktop, moving them beyond the data center.
Where Muse Spark Open fits into the broader Meta and Apple landscape
Muse Spark Open is a new open-weight AI model from Meta. The company describes it as the most capable open-weight model ever created in the United States. Its main selling point is that its weights are available for others to use and develop further.
Apple supports both MLX, a framework for machine learning on Apple Silicon, and the Mac Studio hardware. This collaboration reflects Apple’s desire to promote on-device AI, while giving Meta a way to demonstrate Muse Spark Open’s capabilities outside the data center.
Based on the available information, how much further has Muse Spark Open advanced?
The source confirms several capabilities of Muse Spark Open, but provides no previous model for a direct comparison. It is therefore still impossible to determine how much it has improved.
| Factor | Muse Spark Open | Previous model |
|---|---|---|
| Total parameters | 816 billion | No data |
| Desktop operation | Runs on 2 Mac Studio computers with M5 Ultra chips | No data |
| Capabilities | Multimodal tasks, complex software development, and multi-step reasoning | No data |
| Test results compared with the previous version | No data | No data |
Muse Spark Open’s capabilities in real-world tasks
Muse Spark Open is a multimodal model, making it suitable for tasks that involve handling multiple types of data within a single project, such as analyzing images alongside text or code.
With a total of 816 billion parameters, the model is positioned for large and complex challenges. Its long-horizon capabilities are suited to projects that require continuous work across multiple steps.
For developers, its strengths lie in complex software engineering and multi-step reasoning, making it suitable for coding, problem analysis, and developing solutions that must be built sequentially.
How much can it be compared with other models based on this information?
The source confirms that MLX is an “array framework” for machine learning on Apple Silicon, not a model. It therefore should not be directly compared with Muse Spark Open in terms of performance. Apple’s other two open models are mentioned only as being part of the same technical post, with insufficient names or details to draw any conclusions.
| Factor | Muse Spark Open | MLX | Apple open model 1 | Apple open model 2 |
|---|---|---|---|---|
| Type | AI model | Framework | Open model | Open model |
| Performance information | Capability description available | No direct comparison data | No data | No data |
| Mac operation information | Two M5 Ultra Mac Studios | Runs on Apple Silicon | No data | No data |
For now, it can only be compared in terms of roles and disclosed information. It is still impossible to determine which is more capable than other competitors.
Strengths and limitations apparent from this announcement
Pros
- +Can run on Mac Studio computers using Apple Silicon chips
- +Works with MLX and supports multimodal tasks
- +Suitable for coding and multi-step reasoning
Cons
- −No independent test results are available in the source
- −No details about usage costs
- −No direct comparison data with other models
Costs the news has not yet addressed
- How much do two Mac Studio computers with M5 Ultra chips cost in total?
- How much power does running Muse Spark Open require?
- How are MLX and the model installed, and are there any additional costs?
- How much storage space and system resources are required?
- How much skill or time is needed to configure the system to run the model locally?
- How do the long-term costs of running the model locally compare with using a cloud service?
What does the collaboration between former rivals reflect about on-device AI?
Meta and Apple are often viewed as competitors, but their collaboration over several months, along with the public support expressed by the CEOs of both companies, suggests that on-device AI may require models and hardware that work well together.
The Muse Spark Open case is not yet proof that everyone will be able to run large AI models locally with ease. However, it is a sign that technology companies are beginning to create more room for collaboration. It will be worth watching whether this relationship leads to more accessible on-device AI use.