Superwhisper, which also builds its own dictation software, released s1-mini, a 596-million-parameter model whose entire job is scrubbing raw speech-to-text output into readable prose. It strips filler words, resolves the false starts and self-corrections dictation software leaves behind, adds punctuation and capitalization, and spells out numbers, dates, currency figures and email addresses. The model will not chat, answer questions, or follow open-ended instructions. It performs a single transformation and nothing else.
That narrowness is the point. Superwhisper fine-tuned s1-mini from Alibaba’s Qwen3-0.6B and reports 94.8 percent token accuracy on a self-assembled evaluation covering 7,519 English test cases, a figure the company measured itself rather than one confirmed by an outside benchmark. The quantized build fits in 462 MiB and, according to the model card, is small enough to run smoothly on ordinary laptop hardware, no GPU server required.
The release lands while most model news is about parameter counts climbing into the hundreds of billions. s1-mini runs the opposite direction: a sub-billion-parameter model tuned for exactly one job, sized to sit on a laptop rather than a rack. That is the more interesting story here than the accuracy figure, since the industry’s scaling narrative rarely leaves room for a model this deliberately small, or this willing to refuse everything outside its one task.
Running s1-mini correctly demands more care than a typical inference call. Every prompt needs a fixed system message plus a control line: three bracketed settings covering a formality register that ranges from casual to formal, a structure choice between flowing prose and grouped bullet points, and a destination flag distinguishing plain text from email layout. Superwhisper’s documentation warns that skipping the control line, rewording the fixed system message, or leaving Qwen3’s default reasoning mode switched on produces garbled or blank output. That is a narrower, more fragile integration surface than developers typically expect from a small utility model, and it is worth testing before shipping.
s1-mini carries Apache 2.0 licensing plus one added condition: the model’s name and its maker must appear together, spelled exactly as Superwhisper specifies, in any product built on it. Superwhisper positions the model as an add-on layer, meant to sit inside meeting-note tools, live-captioning products and dictation software rather than ship as a standalone system. English is the only supported language in this v1 release, and the company has not said when or whether that will expand.
Teams currently hand-rolling regex cleanup on top of Whisper or another speech recognizer should treat s1-mini’s CPU footprint and permissive license as a reason to benchmark it against their existing normalization step. If the 94.8 percent figure holds up outside Superwhisper’s own test set, a 0.6 billion parameter model bolted onto an ASR pipeline could replace a layer of custom string handling that most voice products currently maintain by hand.
Superwhisper published the s1-mini model card on Hugging Face on May 14, 2025.
