Superwhisper just dropped its S1 model family, and the interesting one isn’t the cloud stuff. S1-Voice and S1-Language are hosted services — speech-to-text and instruction-based cleanup, fine, whatever. But S1-mini is different: a 0.6B-parameter text normalizer released with open weights on Hugging Face under Apache 2.0. It doesn’t transcribe audio and it doesn’t chat. It sits after your ASR pipeline and rewrites raw transcripts into something a human would actually want to read.
The model is fine-tuned from Qwen3-0.6B with 596M unique parameters, English only in release v1, and handles about 1,000 tokens at a time. It strips filler words, resolves false starts to what the speaker actually meant, applies punctuation and capitalization, and converts spoken formats into written ones — say “support at superwhisper dot com” and you get a proper email address back. The entire interface is a three-axis control line placed above the transcript: Styling (casual through formal), Structure (prose or lists), and Context (general or email). Every combination was trained, which is a nice touch. But stick to the documented values — send something outside those sets and the output can garble.
There are two deployment gotchas that will trip people up. First, you must set enable_thinking=False. Qwen3’s chat template defaults to thinking mode on, and S1-mini was trained with it off. Skip that flag and you’ll get nothing usable. Second, decode greedily. The GGUF builds still carry Qwen3’s inherited temperature of 0.6, so pass temperature 0 explicitly on every request. In llama.cpp, use —jinja with —chat-template-kwargs to disable thinking rather than —reasoning-budget 0, which apparently degrades output quality.
Superwhisper reports 94.8% token accuracy on a held-out set of 7,519 cases across 104 transcripts, measured greedy on the quantized Q4_K_M build. It nails email greetings 99.3% of the time and sign-offs 97.9%, and it correctly produces empty output 98.6% of the time when there’s nothing to transcribe. Those are vendor numbers on an internal test set, so treat them with appropriate skepticism. But the deployment story is genuinely appealing: the Q4_K_M GGUF is a 462 MB file that runs on a laptop CPU. Healthcare, legal, and financial teams can run it behind a VPC where transcripts never leave the network. That’s the real pitch here — a small, focused model that solves one problem well without forcing you to ship audio data to someone else’s cloud.
