Skip to main content
Local LLM Guide

Hardware

AMD Strix Halo 128GB unified memory

Check how much shared memory your graphics processor and model runner can use. Leave room for the operating system, context, and other applications.

Example: 128GB unified memory

Model sizes to check

32B · 70B · some 120B Q4 estimates

Likely bottleneck

Shared memory used by the model, context, operating system, and other applications

Good first workload

Chat, coding, and moderate document workflows

Upgrade trigger

Move to more unified memory when normal applications or longer context force repeated swapping or model downsizing.

What this machine is suited for

Check how much shared memory your graphics processor and model runner can use. Leave room for the operating system, context, and other applications.

  • 32B
  • 70B
  • some 120B Q4 estimates

What will limit it

Unified memory is flexible, but it is not free GPU memory. Every open application and every increase in context uses the same pool as the model.

  • The memory values describe an example configuration; check the actual computer before choosing a model.
  • The estimate changes with the exact model file, context length, runtime, GPU offload, drivers, thermals, and other running apps.
  • A model that technically loads may still be constrained, unstable, or memory-limited for ordinary use. Treat “fits” and “feels good” as different questions.

Before moving to a larger model

  1. Use a common 4-bit build.
  2. Keep context modest until the first run is stable.
  3. Close memory-heavy applications and watch actual memory pressure.
  4. Increase model size only when the smaller model is not solving the task.

What to upgrade

Move to more unified memory when normal applications or longer context force repeated swapping or model downsizing.