Run Llama 3.2 3B on a Mac
Meta's Llama 3.2 3B Instruct — a safe, widely-fine-tuned baseline that many downstream tools assume. 82 tok/s on an M2 Pro in 2.4 GB resident.
3 B dense · 4-bit Llama-ecosystem compatibility · everyday chat
One command
$ rapid-mlx serve llama3-3b-4bit
Weights (1.7 GB) download on first run; you get an OpenAI-compatible
endpoint at http://localhost:8000/v1. No rapid-mlx yet? It's one line —
curl -fsSL https://rapidmlx.com/install.sh | bash — or take the
desktop app.
Measured on real hardware
| Machine | Decode | First token | Peak memory | Cold boot | Weights |
|---|---|---|---|---|---|
| Mac mini M2 Pro · 32 GB | 82.2 tok/s | 0.25 s | 2.4 GB | 3.8 s | 1.7 GB |
Measured on rapid-mlx 0.12.10, 2026-08-11. Decode: median of 3 runs, temperature 0, 256-token saturating generation, engine-reported token counts, unique salt per request (no prefix-cache hits). TTFT: median of 3, short prompt. Boot: process spawn to first completed token. Peak RSS: 0.5 s sampling across the run.
Will it fit your Mac?
Peak resident memory measured 2.4 GB during a 256-token generation. The KV cache grows with context length, so treat that as a floor, not a ceiling. For the conservative install-default placement see the hardware tiers table; to compare against every model your RAM can hold, use the live picker.
Variants & alternatives
qwen3-4b-instruct-2507-4bit— Qwen3 4B Instruct — stronger at a similar size
FAQ
How much memory does Llama 3.2 3B need on a Mac?
Measured peak resident memory was 2.4 GB on rapid-mlx 0.12.10 during a 256-token generation (M2 Pro, 32 GB). The weights are 1.7 GB on disk. Longer contexts grow the KV cache beyond this, so leave headroom.
How fast is Llama 3.2 3B on Apple Silicon?
We measured 82.2 tokens/sec sustained decode and 0.25 s time-to-first-token on a Mac mini M2 Pro (32 GB), median of 3 runs at temperature 0.
How do I run Llama 3.2 3B locally?
Install rapid-mlx (curl -fsSL https://rapidmlx.com/install.sh | bash, or brew install rapid-mlx), then: rapid-mlx serve llama3-3b-4bit — the weights download on first run and you get an OpenAI-compatible endpoint at localhost:8000/v1 that works with Cursor, Claude Code, Aider, and any OpenAI client.