Master Ollama on macOS: Install, Run, and Optimize Large Language Models

1. Installation and Basic Configuration

Ollama can be installed on macOS via two methods. The recommended one‑click script runs: curl -fsSL https://ollama.com/install.sh | sh The script downloads and installs Ollama automatically. Alternatively, download the .dmg from the official website, drag Ollama.app into the Applications folder, and allow execution in System Settings → Privacy & Security if the system blocks the developer.

Verify the installation by opening a terminal and executing: ollama --version A version number (e.g., 0.5.7) confirms success.

Default directories are:

Models: ~/.ollama/models Logs: ~/.ollama/logs Config: ~/.ollama/config.json You can change the model storage location by setting the OLLAMA_MODELS environment variable, e.g.:

# Add to ~/.zshrc
export OLLAMA_MODELS="/Volumes/ExternalSSD/ollama_models"

2. Core Commands and Model Management

Ollama’s CLI mirrors Docker’s style and provides the following essential commands: ollama run <model> – launch a model (e.g., ollama run llama3). ollama pull <model> – pre‑download a model. ollama list – list installed models. ollama show <model> – display model details. ollama rm <model> – delete a model. ollama serve – start the background service (usually auto‑started).

Popular 2026 models include:

llama3:8b – balanced performance.

qwen:7b – strong Chinese language capabilities.

deepseek-coder:6.7b – best for code generation.

phi3:3.8b – lightweight, suitable for M1/M2 base Macs.

Example of running Llama 3:

# First run will download the model
ollama run llama3
# In the interactive prompt, type your question
> 你好，请介绍一下你自己。

3. Advanced Techniques and Deep Usage

3.1 Create a custom model with a Modelfile

Write a Modelfile that defines the base model, system prompt, and parameters:

FROM llama3
SYSTEM """
你是一位专业的技术顾问，回答问题时请保持简洁、准确，并使用 Markdown 格式。
"""
PARAMETER temperature 0.7
PARAMETER num_ctx 4096

Build and run the custom model:

ollama create my-tech-assistant -f ./Modelfile
ollama run my-tech-assistant

3.2 Enable the OpenAI‑compatible API service

Ollama includes an API server listening at http://localhost:11434. Documentation is available at http://localhost:11434/api/docs.

Python example to generate a response:

import requests
response = requests.post(
    'http://localhost:11434/api/generate',
    json={
        'model': 'llama3',
        'prompt': '解释一下量子计算的基本原理。',
        'stream': False
    }
)
print(response.json()['response'])

3.3 Integrate with Open WebUI

Open WebUI provides a ChatGPT‑like web front‑end. Deploy it with Docker:

# One‑click Docker deployment
docker run -d -p 3000:8080 \
  --add-host=host.docker.internal:host-gateway \
  -v open-webui:/app/backend/data \
  --name open-webui --restart always \
  ghcr.io/open-webui/open-webui:main

After the container starts, open http://localhost:3000 and set the Ollama address to http://host.docker.internal:11434.

4. Performance Tuning and Troubleshooting

4.1 Chip‑specific model recommendations

M1/M2 (base) – use phi3 or llama3:8b‑instruct‑q4_K_M with ≥8 GB RAM.

M1 Pro/Max, M2 Pro/Max – llama3:8b or qwen:7b with ≥16 GB RAM.

M3 Max/Ultra – large models like llama3:70b or qwen:72b with ≥32 GB RAM.

For devices with limited memory, prefer quantized variants (e.g., q4_K_M) for smaller size and faster inference.

4.2 Common issues

"command not found: ollama" – the binary is not in PATH. Add /usr/local/bin or restart the terminal.

Slow model download – try a domestic mirror or use a proxy.

Excessive fan noise – the model is too large for the hardware. Switch to a smaller or quantized model, or reduce num_ctx in the Modelfile.

5. Conclusion

Ollama transforms how AI can be used on a personal computer by providing a simple installation path, a Docker‑style CLI for model management, support for custom Modelfiles, an OpenAI‑compatible API, and easy integration with web front‑ends like Open WebUI. Whether you are a developer, researcher, or casual user, you now have the tools to run powerful LLMs locally with full control over privacy and offline availability.