Back to docs

Llanite guide

Local AI agent on Mac

Running a local AI agent on a Mac is straightforward with Apple Silicon hardware. M-series chips use unified memory shared between CPU and GPU, which means large models load efficiently without a separate graphics card. Llanite and Ollama together get you from zero to a running local agent stack in three commands — no manual configuration required.

Why Mac works well for local AI

Apple Silicon Macs use unified memory architecture, where the CPU and GPU share the same memory pool. Ollama takes advantage of this through Metal GPU acceleration, which means model inference runs on the GPU without the overhead of transferring data between separate memory banks. A 16 GB M-series Mac can run capable 7–9 billion parameter models, and a 32 GB Mac can run 27 billion parameter coding models comfortably. This makes Mac one of the most practical platforms for local AI without a dedicated NVIDIA GPU.

Understanding RAM on Apple Silicon

Ollama loads model weights into Metal GPU buffers on macOS. These buffers live in the unified memory pool but don't appear in standard process RAM counters like Activity Monitor's RSS column — this is normal and not a sign of low memory usage. Llanite accounts for this in its hardware fit check and in the /ram command, which lists the loaded model separately under a Metal line. For sizing: 8 GB is enough for the Chat stack (Gemma 3 9B), 16 GB suits the Compact and Lite coding stacks (8–9B models), and 32 GB or more is needed for the Standard stack (27B Qwen 2.5 Coder).

Installing Ollama on Mac

Download Ollama from ollama.com. It installs as a macOS app and adds a menu bar icon. When the app is running, Ollama's local model server is available on port 11434. You can also start it from the terminal.

$ ollama serve$ ollama list

Installing Llanite on Mac

Llanite is distributed via npm and requires Node 20 or later. The easiest way to install Node on Mac is via Homebrew or nvm. Once Node is available, install Llanite globally and run the doctor command to confirm Ollama is reachable.

$ npm install -g @llanite/cli$ llanite doctor

Choosing a stack for your Mac

Llanite checks your system RAM at install time and rates each stack as comfortable, recommended, tight, or incompatible. Browse the catalog, then inspect the stack you want to confirm the fit before installing.

  1. 1Check your RAM: Apple menu → About This Mac → Memory
  2. 2Browse available stacks and their RAM requirements: llanite stacks
  3. 3Inspect a stack before installing: llanite inspect local-coder
  4. 4Run the doctor check for a specific stack: llanite doctor local-coder

Running a local agent session on Mac

Once a stack is installed, llanite run launches the terminal UI agent session. The status line shows the active model, runtime, and agent. The system stats footer shows live RAM and swap pressure, with the Metal GPU buffer reported separately.

$ llanite install local-coder$ llanite run local-coder

FAQ

What Mac specs do I need for a local AI agent?

Any Apple Silicon Mac (M1 or later) can run a local AI agent. 8 GB RAM supports small chat models. 16 GB supports compact coding models like the Lite and Compact stacks. 32 GB or more supports the Standard coding stack with a 27B model.

Does Llanite work on Intel Macs?

Llanite runs on Intel Macs, but Ollama's performance on Intel hardware is significantly lower than on Apple Silicon because there is no Metal GPU acceleration. Model inference runs on the CPU only, which is much slower for larger models.

Why does Activity Monitor show low memory usage for Ollama?

Ollama loads model weights into Metal GPU buffers in unified memory. These buffers don't appear in standard process RSS counts in Activity Monitor. The model is loaded and using memory — it just isn't reflected in the standard view. Llanite's /ram command shows the Metal buffer usage as a separate line.

What is the best stack for a 16 GB M2 Mac?

The Lite stack (lightweight-coder) with a 9B Qwen model is a good choice for 16 GB machines. It supports file read/write, shell, and git tools and fits comfortably in 16 GB of unified memory. The Compact stack (repo-reviewer) is also a good option for read-only code review tasks.

How do I monitor RAM usage during a local agent session?

Type /ram in the TUI during a session to see a breakdown of RAM usage by process, including the Ollama Metal buffer. The system stats footer at the bottom of the TUI also shows live RAM and swap pressure, refreshing every few seconds.

Can I run a local coding agent on a MacBook Air?

Yes. MacBook Air with M1/M2/M3 and 16 GB of unified memory handles the Lite and Compact stacks well. The 8 GB model is limited to the Chat stack. MacBook Air has no active cooling, so sustained heavy model inference may cause thermal throttling on long sessions.