Skip to content

Local LLM Models Guide

Last updated: 2026-01-26 Hardware: MacBook Pro M1 Pro, 32GB RAM

Quick Reference: Models for Your Mac (32GB RAM)

Model Size Speed Best For
qwen3-coder 19GB Fast (MoE, 3.3B active) Agentic coding, best overall
qwen2.5-coder:32b 20GB ~15-20 tok/s High quality code gen
qwen2.5-coder:14b 9GB ~50 tok/s Fast coding tasks
devstral 14GB ~30 tok/s Codebase navigation, file ops

Model Details

ollama pull qwen3-coder
  • Architecture: MoE (Mixture of Experts)
  • Parameters: 30.5B total, only 3.3B activated
  • Size: ~19GB
  • RAM needed: 32GB (perfect for your Mac)
  • Why it's good: Smarter than dense models, runs fast due to MoE
  • Best for: Agentic coding, multi-step tasks, code analysis

qwen2.5-coder:32b

ollama pull qwen2.5-coder:32b
  • Architecture: Dense transformer
  • Parameters: 32.5B
  • Size: ~20GB
  • RAM needed: 24-32GB
  • Benchmark: Best open-source on EvalPlus, LiveCodeBench, BigCodeBench
  • Best for: High quality code generation, complex problems

qwen2.5-coder:14b

ollama pull qwen2.5-coder:14b
  • Parameters: 14B
  • Size: ~9GB
  • Speed: ~50+ tok/s on M1 Pro
  • Best for: Fast iterations, simpler tasks, when speed matters

devstral

ollama pull devstral
  • By: Mistral AI
  • Size: ~14GB
  • Specialty: File system operations, code navigation, large codebases
  • Best for: Codebase exploration, documentation audits, multi-file edits

Benchmark Comparison (SWE-Bench Verified)

Model Score Type
Claude 4 Sonnet 72.7% Closed-source
Claude 4 Opus 72.5% Closed-source
Qwen3-Coder 69.6% Open-source
DeepSeek-V3 ~68% Open-source

Takeaway: Open-source is only ~3% behind Claude now.

Models That WON'T Fit (32GB RAM)

Model Size Issue
qwen3-coder:480b 163-368GB Way too big
deepseek-v3 400GB+ Won't fit
Any 70B+ dense model 40GB+ Will swap, very slow

Usage Patterns

For Code Research/Documentation

ollama run qwen3-coder
# or
ollama run devstral

For Fast Code Generation

ollama run qwen2.5-coder:14b

For Highest Quality (Slower)

ollama run qwen2.5-coder:32b

Running with Context

Pipe a file

cat README.md | ollama run qwen3-coder "Analyze this README and suggest improvements"

Interactive session with system prompt

ollama run qwen3-coder --system "You are a senior software architect auditing a codebase."

Free Cloud Alternatives (When Local Won't Cut It)

Service Models Limit
Gemini CLI Gemini 2.5 Pro 1000 req/day
Groq Llama 3.3 70B Very fast, free tier
Together.ai All major models $25 free credit
OpenRouter Everything Pay-per-use

Check What's Installed

ollama list

Check Available Space

df -h /
du -sh ~/.ollama/models/

Sources