Perplexity’s Local AI Demands 36GB RAM, Leaving Most Mac Users Behind

Perplexity now lets subscribers run its Computer agent entirely on their own Mac — provided that Mac packs an Apple Silicon chip and at least 36GB of unified memory. The requirement immediately excludes every base-model MacBook Air, every 14-inch MacBook Pro with the standard M3 chip, and the entire current Mac mini lineup. Only Max- and Ultra-tier chips, or Pro chips configured with extra memory at purchase, clear the bar.

The company unveiled the local-only mode Tuesday for Pro, Max, and Enterprise tiers. Two models ship out of the box: Qwen 3.6 and PPLX Pro, a Perplexity-tuned variant of the same weights. If a task stalls, the app can still reach out to any of 15-plus cloud models, but code execution stays inside a local sandbox either way.

Apple’s unified memory architecture is the bottleneck. Unlike discrete GPUs that can swap to system RAM, Apple Silicon shares a single pool between CPU and GPU. Large language models need that pool resident for inference speed; 36GB is the floor Perplexity set for the full orchestrator, harness, and model weights to fit without paging. An M3 Max MacBook Pro starts at $3,199 with 36GB. An M2 Ultra Mac Studio begins at $3,999.

The move mirrors a broader shift. Apple’s own Apple Intelligence, Microsoft’s Copilot+ PCs, and open-source projects like Ollama all treat on-device inference as a premium tier gated by memory capacity. Perplexity’s “Hybrid Compute” launch last month already split workloads between cloud and local models; today’s release simply removes the cloud leg for users who can afford the hardware.

Subscribers download the updated Perplexity Mac app, then follow an in-app wizard to pull the selected model weights — several gigabytes each — onto local storage. No separate container runtime or command-line setup is required. The company says the same sandbox security model applies whether the model runs locally or in its data centers.

For the majority of Mac owners, the practical path remains Hybrid Compute: offload heavy reasoning to the cloud, keep lighter tasks and private data on device. Perplexity frames local-first as a privacy and cost-control feature. Whether it becomes a mass-market option depends on Apple’s next silicon generation pushing 36GB into entry-level configurations.