RUNLOCALAIv38
->Will it run?Best GPUCompareTroubleshootStartLearnPulseModelsHardwareToolsBench
Run check
RUNLOCALAI

Independently operated catalog for local-AI hardware and software. Hand-written verdicts. Source-cited claims. Reproducible commands when we have them.

OP·Fredoline Eruo
DIR
  • Models
  • Hardware
  • Tools
  • Benchmarks
TOOLS
  • Will it run?
  • Compare hardware
  • Cost vs cloud
  • Choose my GPU
  • Prompting kits
  • Quick answers
REF
  • All buyer guides
  • Learn local AI
  • Methodology
  • Glossary
  • Errors KB
  • Trust
EDITOR
  • About
  • Author
  • How we make money
  • Editorial policy
  • Contact
LEGAL
  • Privacy
  • Terms
  • Sitemap
MAIL · MONTHLY DIGEST
Get monthly local AI changes
Monthly recap. No spam.
DISCLOSURE

Some links on this site are affiliate links (Amazon Associates and other first-class retailers). When you buy through them, we earn a small commission at no extra cost to you. Affiliate links do not influence our verdicts — there are cards we rate highly that we don't have affiliate relationships with, and cards that sell well that we refuse to recommend. Read more →

© 2026 runlocalai.coIndependently operated
RUNLOCALAI · v38
  1. >
  2. Home
  3. /Learn
  4. /Courses
  5. /Local AI on macOS
  6. /Ch. 8
Local AI on macOS

08. Mac Model Selection Guide

Chapter 8 of 15 · 15 min
KEY INSIGHT

Model selection on Apple Silicon is a RAM calculation first and a capability match second—pick the largest model that fits in your available memory with headroom for the context window.

Model selection on Apple Silicon is not just about capability—it is about fitting within your RAM and achieving acceptable throughput. The two variables are model size (parameters × quantization) and hardware (chip generation and RAM).

A practical selection table for Apple Silicon:

Model Parameters Quantization RAM Needed Target Chip
Phi-3.5-mini 3.8B Q4_K_M 2.4 GB M1+
Llama 3.2 1B 1B Q4_K_M 0.7 GB M1+
Llama 3.2 3B 3B Q4_K_M 1.9 GB M1+
Llama 3.2 7B 7B Q4_K_M 4.9 GB M1 Pro+
Mistral 7B 7B Q4_K_M 4.9 GB M1 Pro+
Llama 3.1 8B 8B Q4_K_M 5.4 GB M2 Pro+
Qwen 2.5 14B 14B Q4_K_M 8.8 GB M2 Pro 24GB+
Llama 3.1 70B 70B Q4_K_M 43 GB M2 Ultra 128GB

The 8 GB machines (M1/M2 base) are viable for 1B–3B models. They can load a 7B model but the memory headroom is too tight for comfortable use at longer context lengths.

The 16 GB machines (M1/M2 Pro base, M3 base) handle 7B models well at short contexts and 3B models at long contexts. This is the sweet spot for most developers.

The 24 GB+ machines (M3 Pro, M3 Max, M4 Pro) run 13B–14B models comfortably and can push into 30B territory with careful quantization selection.

Ollama model search:

# Search available models
ollama list
ollama search llama          # requires internet, lists from registry

# Pull the model
ollama pull llama3.2:3b

MLX model selection via the community hub:

# Find model size and recommended hardware
from mlx_lm import ls
models = ls()
# Returns list of model names with sizes
# mlx-community/Llama-3.2-3B-Instruct-4bit ~1.8GB
EXERCISE

Identify three models from Ollama that fit in your available RAM using the Q4_K_M calculation from Chapter 7. Pull the smallest one and run it. Note the time-to-first-token and tokens per second.

← Chapter 7
Unified Memory Explained
Chapter 9 →
Performance Tuning