Min VRAM needed
1.4 GB
Weights: 0.4 GB
KV cache: 0.4 GB
Overhead: 0.5 GB
The short answer
For a new Mac dedicated to local agents, the 48 GB or 64 GB M5 Pro is the practical Mac mini target. The M6 family tops out at 32 GB, so it is better suited to smaller tested profiles than long-context 27B agent workloads. Apple lists 307 GB/s memory bandwidth and up to 64 GB unified memory for M5 Pro. If you need more memory or bandwidth, move to Mac Studio M5 Max or M5 Ultra rather than planning an eGPU.
Get more guides like this in your inbox
No spam. Unsubscribe anytime.
Critical correction: no Apple-silicon eGPU
Apple's current eGPU support documentation requires an Intel processor. M-series Mac mini and Mac Studio systems cannot use a Thunderbolt NVIDIA or AMD eGPU for local inference. External-GPU expansion is a Windows/Linux mini-PC strategy; the Mac strategy is to buy enough unified memory at purchase time.
Our Mac test system
The isolated test ran on a Mac Studio with M4 Max, 36 GB unified memory, arm64, and macOS 26.4.1. It did not use the host's installed Ollama state. The test downloaded Ollama 0.32.14's signed app, verified SHA-256, strict code signing, and Gatekeeper, then used a private model cache and non-default port. That is close evidence for Apple silicon, not a claim that every Mac mini has identical performance.
Install the verified runtime
Follow the Apple-silicon section of the maintained Ollama install guide. It uses the official Ollama-darwin.zip archive with SHA-256 72f8545a… and verifies codesign plus Gatekeeper before execution. Avoid the standalone tar path for this tutorial because the app bundle gives a verifiable Apple signing boundary.
Mac memory tiers
16 GB: small chat models and experimentation. 24/32/36 GB: strong 9B use and some 20–27B direct inference, but 64K agent context can be tight or slow. 48/64 GB: recommended Mac mini M5 Pro agent tier. 64–128 GB Mac Studio M5 Max: larger quantized models and more context headroom. The M5 Ultra starts at 96 GB and scales beyond the M5 Max tier. Model file size is not the whole requirement; KV cache, runtime buffers, the OS, and tools also consume unified memory.
Qwen 3.8 27B MLX result
The official Ollama profile contained 1,210 objects totaling 18,174,721,847 bytes. Every object matched its manifest SHA-256. A cold direct request returned the exact marker in 14.9 seconds, including a 5.7-second load. This proves the signed runtime and model worked on the named 36 GB M4 Max system.
Why the Hermes profile is still experimental
Hermes 0.20.1 installed from exact commit f80f453a… and its frozen lockfile on the same Mac. With Ollama and Hermes both configured for 64,000 tokens, the one-shot Hermes marker did not complete within seven minutes. The direct model was healthy, so the failure is specifically the full harness/context workload. We will not label Qwen 3.8 27B MLX a recommended Mac Hermes profile until that acceptance test passes with a tuned configuration.
What to deploy today
Use Qwen 3.5 9B as the conservative Hermes starting profile and Qwen 3.8 27B MLX as a direct-inference/experimentation profile on 32/36 GB or larger Macs. For a new purchase, prefer 48/64 GB M5 Pro when the Mac will be a long-lived local-agent appliance; move to Mac Studio when the tested profile needs more memory. Re-run the Hermes exact-marker test after every model, runtime, context, or harness change.