Since this got some interest, I decided to share a simple example you can run on your computer:

https://github.com/AdamISZ/meatier

The example here is with a 3B model, needing about 12GB in RAM (?) and only takes a few seconds to encode/decode:

(venv)% python runme.py
Using device: mps
Model alias: qwen-3b
Loading local model from ./models/Qwen--Qwen2.5-3B
Loading checkpoint shards: 100%|██████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 32.09it/s]
Plaintext : b'Attack at dawn!'
Encoding...
Encoding: 152/152 bits, 82 tokens generated...
Stegotext : In 1492, when Christopher Columbus landed on the New World, he thought he had reached India. But Indian people had already lived in the Americas for thousands of years with little contact with the outside world. They knew about the land and waters, unlike Columbus who was ignorant about this particular region.
While Columbus' expedition of 1492-1493 can be considered America's "first landing," that is
Decoded 82 bits so far...
Recovered : b'Attack at dawn!'
Success: round-trip verified.

That's generous though, most of the other examples looked a lot more stupid.

Also notice, the hidden text:stegotext ratio is 1:20, or so. While that's bad, it is viable for verbal communication, and most types of steganography are a lot worse.

You can change the secret key, the hidden message and the initial prompt to play with it.