MACNATIVE — LOCAL AI TEST WORKSHEET
A blank worksheet for your own observations. No benchmark results are supplied.
Guide: https://macnative.io/blog/local-ai-slow-mac
Memory guide: https://macnative.io/blog/how-much-ram-local-ai-mac

SETUP
Date:
Mac model / chip / unified memory:
macOS version:
App and runtime versions:
Model name, exact tag, format, and quantization:
Local or remote provider:
Context setting:
Thinking/reasoning mode, if applicable:
Other open apps:
Power source / power mode:
Model location (internal or external disk):

SHORT PROMPT (keep identical between comparable runs)
Rewrite this sentence as a clear, friendly email of no more than
50 words. Keep the dates unchanged:
We received the files on Monday and will send our review on Thursday.
Please confirm whether you want comments on the mobile design too.

RUN 1 — FIRST LOAD, FRESH CHAT
Time to first visible answer:
Time to completed answer:
Memory Pressure before / during:
Swap Used before / during (record change, not just the final value):
Dates and requested confirmation preserved? Yes / No:
Is typing and app switching still comfortable?

RUN 2 — SAME MODEL ALREADY LOADED, FRESH CHAT
Repeat the same fields:

RUN 3 — YOUR REAL TASK
Input length / attachment and question:
Repeat the same fields:

ONE CHANGE TO TRY
Change only one: model size, context, input length, or overlapping workload.
What changed:
What stayed fixed:
Result:
Answer quality and editing needed:

OPTIONAL OLLAMA API TIMINGS
load_duration (ns):
prompt_eval_duration (ns):
eval_count:
eval_duration (ns):
Tokens/second = eval_count / (eval_duration / 1,000,000,000)
Do not compute if duration is zero or absent. This excludes loading and
prompt processing. It does not measure answer correctness.

CONCLUSION
Which stage is slow?
What evidence supports that?
Next step: