Superwhisper has released three models: S1-Voice and S1-Language, which run in the cloud, and S1-mini, which does not.
S1-mini does one job. It takes raw output from a speech-to-text system and turns it into text you could publish. It strips fillers, resolves self-corrections, adds punctuation and capitalisation, and writes spoken numbers, dates and addresses properly.
The numbers
It has 0.6B parameters — 596M unique, 0.44B excluding embeddings — across 28 layers. It is fine-tuned from Qwen3-0.6B and handles English only in this first version.
The quantised Q4_K_M GGUF build is 462MB. On that build, Superwhisper reports 94.8% token accuracy across a held-out set of 7,519 cases, with a text-edit error rate of 11.6%.
The licence is Apache 2.0 with a naming clause. Weights are on Hugging Face.
Why the size matters
At 462MB this runs on a laptop or a phone. That changes what you can build. Dictation apps, meeting notes, live captions and clinical documentation can all clean up their transcripts without sending audio or text off the network.
For anyone working with health or legal recordings, that is the difference between a workable tool and a compliance problem. The gap between "the recogniser heard it" and "this is fit to read" has usually meant a cloud round-trip. Now it does not have to.
Source: MarkTechPost
