Back to blog
TutorialOllamaLLMCloud HostingAPILlamaMistral

Host Ollama in the Cloud on Dedicated Apple Silicon

March 25, 20267 min readby Macyou Team

The short answer: the easiest way to host Ollama in the cloud is a dedicated Apple Silicon machine — an M4 Mac mini from $99/mo runs any 7B–14B model around the clock, and a 64 GB M4 Pro handles 70B-class models. You get Ollama pre-installed, an OpenAI-compatible API endpoint, and no per-token fees. Below is why this beats both a laptop and a GPU instance, and exactly how to set it up.

Why host Ollama in the cloud at all?

Running ollama run llama3.2 on your laptop is great for experimentation, but production breaks it quickly: your app needs the endpoint at 3am when the laptop is asleep; an agent loop needs to poll a model 24/7; your team wants one shared endpoint instead of four local installs. Cloud-hosting Ollama fixes all three — the question is what hardware to host it on.

Why Apple Silicon instead of a GPU instance

Ollama is llama.cpp underneath, and llama.cpp is exceptionally good on Apple Silicon: Metal acceleration plus unified memory means the whole model lives in one memory pool — no VRAM ceiling, no CPU↔GPU copies. The practical consequences:

  • Fixed cost. A dedicated Mac is a flat monthly price. A cloud GPU that runs 24/7 bills by the hour — an A10G-class instance left on all month costs several times more than a Mac mini that serves the same 8B–70B Ollama workloads.
  • Memory, not VRAM. 64 GB of unified memory runs a Q4 70B model. Matching that in GPU VRAM means an A100-class card at datacenter prices.
  • No cold starts. Serverless GPU endpoints spin down and make your first request wait. A dedicated machine keeps the model warm (set OLLAMA_KEEP_ALIVE) permanently.

Setup on Macyou (about 5 minutes)

Macyou deployments ship with Ollama pre-installed and wired to an OpenAI-compatible endpoint, so there is nothing to configure:

  1. Build a Mac — pick a chip and memory for your model size (see the sizing table below).
  2. Pick a model stack from the catalog (Llama, Qwen, Mistral, DeepSeek) — or a clean machine if you want to pull models yourself.
  3. Point your code at the endpoint — it speaks the OpenAI API format:
from openai import OpenAI

client = OpenAI(
    api_key="mcy_live_...",
    base_url="https://dep-xxxx.macyou.cloud/v1",
)
resp = client.chat.completions.create(
    model="llama3.2",
    messages=[{"role": "user", "content": "Hello!"}],
)

You also get SSH, so the normal Ollama CLI workflow works exactly as on your laptop:

$ ssh your-mac
$ ollama pull qwen2.5:14b
$ ollama list
NAME              SIZE    MODIFIED
llama3.2:latest   4.7 GB  2 hours ago
qwen2.5:14b       9.0 GB  just now

Sizing: which Mac for which models

  • M4, 16 GB (from $99/mo) — one 7B–8B model (Llama 3.x 8B, Mistral 7B, Qwen 7B) with headroom, or a 14B on its own.
  • M4, 32 GB — a 32B Q4 model (Qwen 2.5 32B), or two 8B models kept warm side by side.
  • M4 Pro, 64 GB (from $199/mo) — 70B-class at Q4 (Llama 3.3 70B, Qwen 72B), the sweet spot for serious private inference.
  • M4 Max / M3 Ultra (96–256 GB) — Q8 and FP16 builds, multi-model serving, frontier-scale experiments.

Per-model detail — exact RAM by quantization, chip fit, and cost — is in our model hardware guides.

Multiple models on one machine

Ollama hot-swaps models on demand: a request names a model, Ollama loads it, serves, and keeps it warm. On a 32 GB machine two 8B models coexist comfortably; unified memory makes reloads a matter of seconds. For routing across many models behind one API key, the catalog also ships a LiteLLM proxy stack.

FAQ

How much does it cost to host Ollama in the cloud?

On Macyou: from $99/mo for an M4 Mac mini (7B–14B models) to $199/mo for an M4 Pro 64 GB (70B-class), flat, with no per-token or per-hour charges. Annual billing takes 20% off.

Is the endpoint OpenAI-compatible?

Yes — every deployment exposes /v1/chat/completions, /v1/embeddings, and /v1/models. Existing OpenAI SDK code works after changing base_url and the key.

Is my data private?

The model runs on a physical machine dedicated to you — prompts and outputs never leave it for any third-party API. Disks are encrypted and wiped between tenants.

Ready to try? Build a Mac and have an Ollama endpoint running in about 5 minutes.

All posts