Axiom

Appearance

Core ML Compression ​

How to make a Core ML model smaller and faster on device, and the decision that matters most: compressing after training, or training with compression in the loop.

When to Use ​

Use this skill when you're:

  • Fitting a model that's too large for memory or download size
  • Choosing among quantization, palettization, and pruning
  • Deciding between post-training compression and training-time compression
  • Picking a bit width, and knowing which ones need an iOS 18 deployment target
  • Matching a compression method to the Neural Engine or a Mac GPU
  • Recovering accuracy that compression cost you

Example Prompts ​

Questions you can ask your assistant that will draw from this skill:

  • "My model is 4 GB. How do I get it small enough for iPhone?"
  • "Should I quantize or palettize my model?"
  • "My model lost accuracy after 4-bit compression. What now?"
  • "Which compression works best on the Neural Engine?"
  • "Why does coremltools say this compression needs iOS 18?"

What This Skill Provides ​

  • The core decision – When post-training compression is enough and when training-time compression is worth the effort
  • The three families – Quantization, palettization, and pruning, and what each trades away
  • Deployment target and hardware – Which methods need iOS 18, and which suit the Neural Engine or a Mac GPU
  • A bit-width ladder – Stepping down from 16-bit through 8, 6, and 4 bits while measuring each step
  • Re-measuring – Checking accuracy after every compression step instead of assuming it held

Documentation Scope ​

This page documents the coreml-compression skill (in the axiom-ai suite), which your assistant loads when you shrink a Core ML model. If accuracy drops after compression and you need to find which layer caused it, see Core ML Diagnostics.

Resources ​

WWDC: 2024-10159

Docs: /coreml

Released under the MIT License

Copyright © 2026 Charles Wiltgen • v27.6.0