Small models, narrow tasks. Trained from scratch, under 100M parameters, each doing one thing.
The 91M Bulgarian model against INSAIT's BgGPT-Gemma-2-2.6B, 28× larger:
| ours | theirs | |
|---|---|---|
| fluency judging, 2000 pairs | 0.9305 | 0.9175 |
| bits per character, same text | 1.3041 | 1.1600 |
One win, one loss, both on the card. bg-eval re-runs them without trusting us.