Running powerful AI tools on a Mac is becoming less of a cloud-only idea. Perplexity now lets users run its entire Computer experience locally, while Meta has introduced an 800-billion-parameter model designed for Apple Silicon.
These updates point to two different paths for local AI. Perplexity focuses on running workflows directly on a user’s machine, with a lower hardware requirement of 36GB of RAM. Meta’s Muse Spark Open aims at a much larger scale, requiring 1TB of RAM across Mac Studio systems.
Perplexity brings its full Computer experience to Mac
Perplexity introduced Hybrid Compute in September, giving users a way to split AI workloads between frontier models and models running locally. The company has now taken that approach further by allowing the entire Computer experience to run on a user’s own Mac.
That includes the harness and orchestrator, the parts of the system that manage and coordinate a workflow. Perplexity’s local models include Qwen 3.6 and PPLX Pro, and the software can still authorize calls to more than 15 models when a workflow needs extra capability.
Users do not have to choose between local processing and every other model available in the system. The setup can run local models on Apple Silicon while calling on additional models for demanding workflows, keeping those options inside the same Computer experience.
Code and tool execution take place in an isolated sandbox environment. That gives Perplexity’s workflow system a separate place to run those actions, rather than treating code or tools as part of the normal desktop environment.
Perplexity Pro, Max, and Enterprise subscribers can begin using Computer locally on the date of the release. Perplexity Computer can also run fully offline on a Mac, but the machine needs at least 36GB of RAM.
That requirement sets a clear line for who can use the feature. A Mac that meets the memory requirement can run Qwen 3.6 or PPLX Pro directly through Apple Silicon, while machines with less than 36GB of RAM do not meet the stated requirement for the local Computer experience.
“As more demanding workflows become practical to run locally, the machines people already own will take on a larger share of their AI work,” Perplexity said.
Meta pushes local AI toward a much bigger scale
Meta’s Muse Spark Open takes the local model idea to a different level. The model has 800 billion parameters, was optimized for Apple Silicon, and requires 1TB of RAM to run locally.
Meta says Muse Spark Open outperforms the Chinese models GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning. The model is also optimized for Mac Studio systems with M5 Ultra chips, where Apple’s unified memory architecture lets the GPU use up to 512GB per Mac Studio.
Two Mac Studios with M5 Ultra can hold the entire Muse Spark Open model. In that setup, the two computers behave like one supercomputer with 1TB of unified memory, matching the model’s stated local memory requirement.
The hardware cost makes the scale of this setup easy to understand. Running Muse Spark Open on four 256GB M5 Ultra Mac Studios costs roughly $38,000.
Apple has not announced specific pricing or a ship date for the 512GB Mac Studio, but it will be available in late October. That system would offer 512GB of GPU memory per Mac Studio, giving local model users another hardware option for handling large workloads.
Two versions of the local AI future
Perplexity and Meta are presenting different versions of what local AI can mean on a Mac. Perplexity’s approach targets workflows that can run on a machine with 36GB of RAM, combining local models with access to more than 15 models when needed.
Muse Spark Open targets a much larger machine setup. Its 800-billion-parameter design needs 1TB of RAM, and the model’s hardware plan depends on connecting Mac Studios with M5 Ultra chips and using their unified memory.
The difference matters because “running AI locally” does not describe one fixed level of hardware. In Perplexity’s case, local operation includes the full Computer experience, offline use, local models, and sandboxed code or tool execution. In Meta’s case, local operation can mean assembling a 1TB memory system from multiple high-end Macs.
Muse Spark Open weights are releasing today and are available to download from Hugging Face. Mark Zuckerberg said, “Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple.”
Together, the two updates give Mac users a wider range of local AI options. One starts with a 36GB memory requirement for an offline AI workflow. The other reaches 800 billion parameters and needs a terabyte of memory, showing how far local AI can stretch when Apple Silicon and unified memory are combined.
Based on
