Meet S1-mini: Superwhisper’s 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text

Superwhisper has split its new S1 family into three very different jobs. S1-Voice handles cloud speech-to-text. S1-Language does cloud cleanup and formatting. And S1-mini is the one with open weights, a small text normalizer meant to sit after automatic speech recognition and turn rough transcripts into something people can read.

That distinction matters. S1-mini is not a transcriber and it is not a chat model. It rewrites raw ASR output by stripping filler words, settling false starts and self-corrections on the version the speaker actually landed on, and adding punctuation, capitalization, and written forms for numbers, dates, currency, and email addresses. Superwhisper says the model is fine-tuned from Qwen/Qwen3-0.6B and covers English only in release v1.

The open-weights release is the part that makes this interesting outside Superwhisper’s own app. The Q4_K_M GGUF build is 462 MB and can run on a laptop CPU. The model is published on Hugging Face under Apache 2.0 with a naming clause, which means solo developers can bundle it into desktop software while enterprises can keep transcripts inside a VPC. Superwhisper also points to uses in healthcare, legal work, finance, customer support, accessibility, live captioning, dictation tools, meeting notes, and voice-driven editors.

The interface is deliberately narrow. S1-mini takes a fixed system prompt, then a control line with three axes: Styling, Structure, and Context. Styling can be casual, semi-casual, semi-formal, or formal. Structure can be prose or lists. Context can be general or email. Superwhisper says every combination was trained, but values outside those sets can degrade output. The model also refuses to add facts you did not say, and filler-only input is supposed to return an empty string.

There are a couple of integration traps. The model was trained with thinking off, so enable_thinking=False is required. It also decodes greedily, so temperature needs to be set to 0 even though the GGUF metadata still carries Qwen3’s inherited sampling values. On its internal held-out set of 7,519 cases, Superwhisper reports 94.8% token accuracy on the Q4_K_M build, 11.6% text-edit error rate, and fewer than 1% looping or truncation failures.