YOUR OWN AI WORKSHOP
Compare the builds.
Inspect the difference.
By Alvin
Each row is one generated function from our 128GB M5 Max. Checks counts four test examples; elapsed seconds includes model loading.
| MODEL / BRIEF | CHECKS | ELAPSED s |
|---|---|---|
| Qwen3.5 4B / notebook | 4/4 | 34.20 |
| Qwen3.5 4B / triage | 4/4 | 17.07 |
| Llama3.2 3B / notebook | 1/4 | 9.70 |
| Llama3.2 3B / triage | 1/4 | 10.87 |
Qwen passed both sets. Llama passed only the empty-list examples: it dropped unconfirmed notebook notes and lost source labels from triage.
Give the field a clear job
if 'confirmed' not in note or note['confirmed']:
report.append({'source_id': source_id,
'quote': quote})
Our notebook fixture expects every note. Llama treated confirmed as a selection rule. The next page adds an explicit inclusion sentence and follows the resulting change.
Decode rate measures reply generation after prompt processing: Qwen produced about 28–29 tokens/s; Llama 28–34. Elapsed times also include loading. We left the operating system’s file cache in place.