SuperWhisper launches S1-mini, a 596M-parameter transcription model

SuperWhisper unveiled its S1 family: cloud models S1-Voice and S1-Language plus the open-weight S1-mini, a 596M-parameter model that runs fully offline on a laptop CPU. S1-mini reaches 94.8% token accuracy and an 11.6% text-edit error rate on 7,519 held-out cases.

• S1-mini has 596M unique parameters with open weights on Hugging Face
• 94.8% token accuracy and 11.6% text-edit error rate on 7,519 cases
• S1-Voice posts 6.8% average WER and 2.2% on LibriSpeech
• Requires enable_thinking=False or it silently returns no output