0.8B local model fine-tuned to match frontier model in dictation cleanup

MoodOdd9657 · reddit · 2026-08-30

The author released SpeakoFlow Mini, an Apache 2.0 fine-tune of Qwen2.5-0.8B specialized for dictation cleanup. Under a fixed short prompt with no reasoning budget, the model scored 70.7% on a specialized benchmark, statistically tying with GPT-4o (Luna) at 65.0%. Fine-tuning improved the base score by 23.4 points. The model is 833MB and runs fully offline.

Original post →

More from Models

Models channel →