There's a little chunk of (headerless) compressed data here:
https://github.com/robinsonb5/EightThir ... 4/Firmware
If it uncompresses to a brief message then some Lorem Ipsum text...

Shaving off even a further two bytes is still pretty cool!

That 68k code is pretty impressive.

174 bytes. I found a LZ4 decompression routine for RISCV and ported it to Thor. It turned out to be 174 bytes in size, but I blew 16 bytes saving and restoring the registers that were used. Still the 68k code was much more compact. I wonder if that is generally the case for a wide variety of apps. I have not tested it yet.

...and sometimes you win the staring competition with the abyss.
I haven't gotten as far as testing it yet, but after waving my fist at it for most of Saturday I am pretty confident I got it down to 66 bytes. That is still with a 2-byte penalty for the format being small-endian and a 4-byte penalty for not using 68020+.

Optimizing for size can be detrimental to optimizing for speed; with more common code executed you can be sure to have known (beneficial) states.