Meta releases 800B parameter Muse Spark Open AI model for Apple Silicon Macs
Meta has open-sourced its largest US model to date, optimized for Apple Silicon but requiring a terabyte of RAM to run locally.
Key points
- Meta released Muse Spark Open, an 800-billion parameter model, as an open-weight release.
- The model is optimized for Apple Silicon and outperforms specific Chinese models on agentic tasks.
- Running the model locally requires two M5 Ultra Mac Studios with 512GB of RAM each.
Meta has released Muse Spark Open, an open-weight artificial intelligence model containing 800 billion parameters, according to a report by Engadget. The company describes this release as the most capable US open-weight model currently available. Meta states that the model is optimized specifically for Apple Silicon hardware, marking a significant shift in how large language models are deployed on consumer-grade workstations.
The release coincides with Meta’s stated policy that the United States should prevent Chinese laboratories from leading in open-source AI development. Benchmarks provided by Meta indicate that Muse Spark Open surpasses two leading Chinese models, GLM 5.3 and Kimi K3, in performance. These benchmarks focus on agentic tasks and visual reasoning capabilities, areas where Meta claims its new model holds a distinct advantage over competitors.
Hardware Requirements and Architecture
Running Muse Spark Open locally demands substantial computational resources that exceed standard consumer configurations. Meta explains that the model requires one terabyte of RAM to operate effectively. To achieve this, users must utilize Apple’s M5 Ultra-equipped Mac Studio machines. Apple’s unified memory architecture allows the M5 Ultra’s integrated GPU to access up to 512 GB of memory per Mac Studio unit.
According to Meta, connecting two Mac Studios enables them to behave like a single supercomputer with a combined terabyte of unified memory. This setup allows the entire model to reside in memory, facilitating local inference without relying on external cloud servers. The company collaborated with developer John Ternus and Apple to optimize the model for this specific hardware configuration, ensuring efficient data handling across the distributed memory architecture.
Availability and Cost Implications
Muse Spark Open weights are available for download immediately from Hugging Face, making the model accessible to developers and researchers today. However, the hardware requirements create a significant financial barrier for most individual users. Apple has not yet announced specific pricing or a ship date for the 512GB Mac Studio, only stating it will be available in late October.
A functional setup requires either two 512GB Mac Studios or four 256GB M5 Ultra Mac Studios. The latter configuration is estimated to cost roughly $38,000. This price point places the technology out of reach for typical independent developers, often referred to as vibe coders. However, larger organizations that have exhausted their token budgets for cloud-based AI services may find this local investment more economical in the long run, provided the model performs as Meta suggests.
What should you do?
Organizations evaluating local AI deployment should assess their current hardware inventory against the 800-billion parameter requirement. If your infrastructure relies on cloud token budgets, calculate the total cost of ownership for local Mac Studio clusters versus ongoing cloud expenses. Monitor Apple’s official announcements in late October for confirmed pricing and availability of the 512GB M5 Ultra Mac Studio units.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Meta
- Audit your current AI inference costs to determine if local hardware deployment offers financial savings over cloud token usage.
- Verify that your organization’s hardware can support the unified memory architecture required for 800-billion parameter models.
- Monitor official channels for confirmed shipping dates and pricing of Apple’s M5 Ultra Mac Studio configurations.
- Review the performance benchmarks provided by Meta to ensure the model meets your specific agentic and visual reasoning needs.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How many parameters does the new Muse Spark Open model contain?
The Muse Spark Open model contains 800 billion parameters, according to Meta’s release announcement.
What hardware is required to run Muse Spark Open locally?
Running the model locally requires two M5 Ultra Mac Studios with 512GB of RAM each, or four 256GB units, totaling one terabyte of RAM.
Is Muse Spark Open available for download now?
Yes, the weights for Muse Spark Open are available to download immediately from Hugging Face.
