What reasoning level in test 2?
Seemed to default to extra-high on Unsloth for me with Unsloth's Q4. (And yeah, it absolutely overran on a test I did.)
I just had "model default" selected; this was with unsloth Q4 also so I assume it's the same. That explains a fair bit about why it was so damn chatty generating the python. :)
Running it on low was better. It wanted to use tools which probably would have been good but couldn't get the in-browser js stuff to work on the image (perhaps it doesn't). It output a completely wrong answer (1979 dime instead of 1967 half-dollar) but at least it didn't loop.
test 1: write a python program to sort a csv by lastname and print out fields. Result: Nailed it. Nice python program. Nice explanation of output, if a little chatty (it gave an example output, which I didn't ask for).
test 2: my standard "identify the type and year of this coin" test. Failed miserably, I gave up after 12,000 tokens of it babbling incessantly about possible misinterpretations of a silver half-dollar. Unsurprising - most small models suck at this test - but slightly annoying it seemed to get in a loop.
Pretty good performance and result for the (very simple) code generation problem, though. One of the first models that seems OK running side-by-side with a browser on my older MBP.