Key points
- Apple researchers demonstrated running Meta’s Muse Spark Open on a cluster of four Mac Studios.
- The setup achieves over 36 tokens per second, exceeding average human reading speed.
- A new MLX version supports local large language models via Apple silicon’s unified memory.
Apple has outlined a strategy for delivering artificial intelligence to Mac users through on-premises infrastructure rather than relying solely on cloud services. According to Computerworld, the company showcased how organizations can use Mac hardware to host large language models locally. This approach targets enterprises that require data privacy and control over their AI workloads. The demonstration highlights Apple’s focus on cost-effective and responsible AI deployment for business environments.
By the numbers
- 36 tokens per second: the output speed of Muse Spark Open on four Mac Studios.
- 4 Mac Studios: the hardware cluster used to run the frontier-class model locally.
- 1 trillion parameters: the approximate size of the Muse Spark Open model being tested.
What the figures mean
The ability to generate over 36 tokens per second on a small cluster suggests that local AI inference is becoming viable for real-time business applications. According to Computerworld, this speed exceeds average human reading rates, making the output suitable for interactive tasks. By using four Mac Studios, organizations can host a frontier-class model with nearly a trillion parameters without sending data to external servers. This keeps sensitive information on-site while leveraging Apple silicon’s unified memory architecture for efficient processing.
View the original post on X
View the original post on X
What should you do?
Managed service providers should evaluate whether their clients have the network infrastructure to support local AI clusters. Teams usually assess storage requirements for large model weights before deploying on-premises solutions. Check whether existing Mac hardware meets the memory specifications for running mixture of experts architectures. Verify that internal security policies allow for local model execution and data handling. Ensure that staff are trained to manage the maintenance and updates of these distributed inference systems.
Operational impact
Organizations must determine if local AI processing aligns with their data residency requirements. IT leaders should assess the total cost of ownership for maintaining Mac Studio clusters versus cloud subscriptions. Procurement teams need to verify hardware availability and supply chain stability for high-end Apple silicon devices. Security officers must review access controls for on-premises models to prevent unauthorized data extraction. This shift requires coordination between infrastructure, security, and application development groups to ensure seamless integration.
Background: Large language models
The difference between large language models and small language models lies in their parameter counts, training data volumes, and memory footprints. Large models handle complex, open-ended reasoning across many domains, while small models execute narrow tasks efficiently on local hardware with predictable latency and lower resource overhead.
What to do and how to stay safe: Apple
- Review current AI data handling policies to see if local processing offers better compliance.
- Audit existing Mac hardware to determine if it supports the latest MLX framework updates.
- Monitor vendor announcements for official support of enterprise-grade AI models on Apple silicon.
- Test local inference performance with sample datasets to benchmark against current cloud solutions.
Step-by-step guide: Shadow AI Explained: 8 Questions Answered for Secure Operations
General security guidance from the BitShield News newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Can any Mac run Meta’s Muse Spark Open?
The demonstration used a cluster of four Mac Studios, indicating that high-end hardware is required for this specific workload.
What is MLX?
MLX is a machine learning framework from Apple that enables researchers to run large language models locally on Apple silicon devices.
How fast is the local AI output?
The cluster generated over 36 tokens per second, which is faster than average human reading speed.
