macOS 26.4.1 VMs on Apple silicon run at 98% single-core CPU speed and are usable with as little as 2 vCPUs and 4 GB RAM.
Key Takeaways
On a Mac mini M4 Pro, a 5-core/16 GB guest hits 98% of host single-core Geekbench 6 score and 95% of GPU Metal score.
The virtual neural engine is the weak point: half-precision and quantised CoreML scores collapse far below host figures.
A 2-core/4 GB VM consumed only 3.1 GB RAM and handled Safari and everyday tasks normally, making MacBook Neo viable as a host.
Minimum practical VM disk is ~60 GB for safe updates; APFS sparse files mean a 100 GB VM uses ~54 GB on disk.
GPU compute (Metal rendering) virtualises well; ANE-dependent AI workloads do not – macOS VMs are not suitable for on-device LLM inference.
Hacker News Comment Review
Commenters on M5 Air note a persistent gap: virtio-gpu passes graphics but not compute GPU, blocking PyTorch GPU acceleration inside any VM or container on macOS.
Discussion questions whether observed memory savings at smaller allocations are caused by core count reduction or simply by the smaller memory ceiling – the relationship is likely the latter.
Notable Comments
@Havoc: PyTorch GPU acceleration remains unavailable in macOS VMs; virtio-gpu only exposes graphics, not compute, making ML isolation a dead end currently.
@JasonHEIN: Flags an open opportunity: no established environment exists for spawning AI agents inside macOS VMs specifically.
@llm_nerd: Clarifies that the weak ANE Geekbench result reflects a test explicitly targeting the neural engine, not a general AI performance ceiling.
