Ollama on Mac 2026: Complete Apple Silicon Setup Guide (M1–M5, Metal GPU)

Quick Setup (3 Commands)

1. 1
   Install Ollama
   Why it matters: `brew install ollama` — one-click installation.
2. 2
   Pull a model
   Why it matters: `ollama pull llama3.1` — downloads Llama 3.1 8B.
3. 3
   Start chatting
   Why it matters: `ollama run llama3.1` — interactive chat interface.

Metal GPU Verification

1. 1
   Run with verbose output
   Why it matters: `ollama run llama3.1:8b --verbose` and look for `ggml_metal_init: found device: Apple M[X]` in the console output.
2. 2
   Check speed during inference
   Why it matters: Observe token generation rate: should be 20–60 tok/s depending on Mac (M5 Pro: ~50 tok/s on Llama 3.1 8B). CPU-only fallback: ~1–5 tok/s.
3. 3
   Monitor GPU utilization
   Why it matters: Open Activity Monitor (Applications → Utilities) and check GPU section. Should show 80-100% GPU utilization during inference if Metal is working.

Model Management

1. 1
   `ollama pull <model>`
   Why it matters: Download model. Example: `ollama pull mistral`.
2. 2
   `ollama list`
   Why it matters: List all downloaded models.
3. 3
   `ollama run <model>`
   Why it matters: Start interactive chat with model.
4. 4
   `ollama rm <model>`
   Why it matters: Delete model to free space.

Memory Optimization for Apple Silicon

• OLLAMA_MAX_LOADED_MODELS: Number of models to keep in memory. Default: 1. Set to 2–3 for multi-model setups.
• GPU layers: By default, Ollama uses all available unified memory. For sub-optimal memory, set `num_gpu_layers` in Modelfile.
• Whisper: Combine with embedding model and LLM — fits in 64GB M5 Pro with Ollama.

Running Multiple Models Simultaneously

export OLLAMA_MAX_LOADED_MODELS=3
export OLLAMA_KEEP_ALIVE=1h
brew services restart ollama

# Now pull all models you need
ollama pull llama3.1:8b
ollama pull llava:7b

# Send requests to each — they stay loaded
curl http://localhost:11434/api/chat -d '{"model": "llama3.1:8b", "messages": [{"role": "user", "content": "Hello"}]}'
curl http://localhost:11434/api/chat -d '{"model": "llava:7b", "messages": [{"role": "user", "content": "Describe this image"}]}'

Auto-Start on Login

# Enable auto-start
brew services start ollama

# Check status
brew services list | grep ollama

# Disable auto-start (optional)
brew services stop ollama

API Setup for Developers

# Chat endpoint (streaming)
curl http://localhost:11434/api/chat -d '{
  "model": "llama3.1:8b",
  "messages": [{"role": "user", "content": "Write a Python function"}],
  "stream": false
}'

# Python example
import requests
response = requests.post(
  "http://localhost:11434/api/chat",
  json={
    "model": "llama3.1:8b",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": False
  }
)
print(response.json()["message"]["content"])

Modelfile Customization

• `ollama create llm-expert -f Modelfile` — builds custom model
• `ollama run llm-expert` — starts interactive chat with your custom model
• `ollama run llm-expert "Code review this function"` — send prompt directly

FROM llama3.1
SYSTEM "You are an expert software engineer reviewing code for security and performance issues. Provide actionable feedback."
PARAMETER temperature 0.7
PARAMETER top_p 0.9

Common Issues and Fixes

• Metal not detected: Verify with `ollama run llama3.1:8b --verbose` and look for `ggml_metal_init: found device: Apple M[X]`. If missing, restart: `brew services restart ollama` or `pkill ollama && ollama serve &`.
• Slow inference (CPU fallback): Cause: Metal failed to initialize, model running on CPU. Check Activity Monitor — GPU usage should be 80-100% during inference. If GPU shows 0%: restart Ollama and check Metal not detected above.
• Out of memory (OOM): Model crashes or response truncates. Cause: model + context + macOS overhead exceeds RAM. Fixes: (1) Use smaller quantization (`ollama pull llama3.1:8b-q4_K_M`), (2) Reduce context (`ollama run llama3.1:8b` then `/set parameter num_ctx 2048`), (3) Use smaller model (`ollama pull phi4-mini` — 2.5 GB).
• Model download stalls: Cause: network throttling or registry/network limits. Fix: `pkill ollama && ollama pull llama3.1:8b` (resumes from previous progress).
• Port 11434 already in use: Another Ollama instance is running or different service uses port. Find: `lsof -i :11434`. Fix: `pkill ollama` then restart.
• Model produces gibberish / random characters: Cause: Modelfile parameters out of range or wrong template. Fix: Pull official model `ollama pull llama3.1:8b` (overwrites custom), then test: `ollama run llama3.1:8b "Hello, how are you?"`.
• Storage filling up: Models stored in `~/.ollama/models/`. Check size: `du -sh ~/.ollama/`. Remove unused: `ollama rm <model-name>`.

Related Articles

• Apple Silicon for Local LLMs: Complete Guide
• MLX vs Ollama vs llama.cpp Comparison
• Best LLM Models for Apple Silicon
• How Much Unified Memory for Local LLMs?
• M5 Pro vs M5 Max Benchmarks
• Mac Mini M5 as Local AI Server
• Apple MLX vs NVIDIA CUDA 2026 — for deciding between Apple Silicon and NVIDIA GPU hardware.