विधि Vidhi

Run AI on your own Mac

Contracts, salary sheets and client files should not have to leave your office to be summarised. In about two hours: check what your Mac can hold, install a free model, ask it about a private document with the Wi-Fi off, and, if you use the Terminal, run Claude Code on it.

What you need

  • A Mac with Apple silicon (M1 or later) on macOS 14 or newer
  • 16 GB of memory at the very least; 24 to 32 GB to run a strong model
  • 20 to 40 GB of free disk space and a good connection for the first download
  • One private document you would not paste into a cloud chatbot

What you will have

  • A free model that runs with the internet switched off
  • Your first private summary of a real document
  • A clear rule for which work stays local and which goes to the cloud
  • Optionally, Claude Code working through a local model

  1. Step 1: Check what your Mac can hold

    5 minutes · About This Mac

    Apple menu, then About This Mac. Note the chip (M1 or later is needed) and the memory. Memory decides everything: a model has to fit in it with room left for macOS. A rough rule: at 4-bit a model needs a little over half a gigabyte for every billion parameters, plus a few gigabytes for the conversation. Keep about a quarter of your memory free for the Mac itself.

  2. Step 2: Pick a model that fits

    10 minutes · LM Studio or Ollama

    With 16 GB, stay with small models of about 8 billion parameters. With 24 to 32 GB, Qwen 3.8 27B at 4-bit (about 15 to 18 GB, Apache 2.0 licence) is the model most people now start with. With 64 GB or more you can run much larger models. On a Mac, choose the MLX build when there is one: it is made for Apple silicon.

  3. Step 3: Install LM Studio and download the model

    30 minutes · LM Studio (free)

    Download LM Studio from lmstudio.ai and open it. Use the search to find the model from the last step, pick the MLX version sized for your memory and download it. When it finishes, load it and open a new chat. The first download is the slow part; after that the model works with no internet at all.

  4. Step 4: Ask it something private, offline

    15 minutes · LM Studio

    Switch the Wi-Fi off. Attach one private document, a contract or a salary sheet, and ask about it. If you get a good answer with the Wi-Fi off, you have proved to yourself that nothing left the Mac.

    You are helping the founder of Sharma Dental Studio. Read the attached document and give me: what it commits us to, every date and amount in it, anything unusual or one-sided, and three questions I should ask before I sign. Quote the clause for each point.

  5. Step 5: Leave the settings alone

    5 minutes · LM Studio

    Keep the model maker’s default settings. Changing temperature and similar sliders is the most common reason a local model starts looping or seems to get dumber. If an answer is poor, change the prompt or the model, not the dials.

  6. Step 6: If you use the Terminal: Ollama and Claude Code

    30 minutes · Ollama · Claude Code

    Install Ollama from ollama.com, then in Terminal type ollama pull qwen3.8:27b to download the model, then ollama launch claude to start Claude Code on it. Ollama’s docs advise a context window of 64k or more for real projects. Claude Code still sends anonymous usage statistics, not your code or prompts, unless you set DISABLE_TELEMETRY=1.

  7. Step 7: Decide what stays local

    10 minutes · A note on your desk

    Write a one-line rule for your team. Private documents, client data, first drafts and large batches of files go to the local model. The hardest analysis, and anything that needs the latest web, goes to the best cloud model. Revisit the rule every quarter: local models are improving every few months.