Superwhisper released the new version. S1 family of modelsS1 – Voice, S1 – Language and S1 – Mini. S1-Voice is a cloud speech-to-text model, and S1-Language is a cloud instruction-following model for cleanup and formatting. It is the one outside of app that has the most interest. S1-miniHugging Face is released in open weights. The S1 Mini is a text normalizer that can handle 0.6B, it’s not a transcriptor or chat software. The S1-mini sits on top of automatic speech recognition to rewrite raw transcripts in clean written format: Fillers and self-corrections are eliminated, punctuation is applied and written numbers, dates and currencies as well as email addresses. This is fine tuned from Qwen/Qwen3-0.6BIn release v1, only English is covered. A three-axis line above the transcription controls all of the functions. Superwhisper reported 94.8% accuracy in tokens on a set of 7,519 held-out cases.
Is it deployable?
No, only the S1-mini. S1-mini Hugging Face has published the software under Apache 2.0 with a naming-clause. S1-Voice, S1 – Language and Superwhisper are all hosted services. They are not self-hosted.
- The Company LevelThe Q4_K_M is available at any level. The Q4_K_M GGUF build It is 462MB and runs on laptops. Solo developers can ship it inside a desktop app. It can be run behind a VPC, where the audio transcripts are not allowed to leave your network.
- IndustriesDeveloper tooling and accessibility, as well as live captioning and healthcare and clinical documentation.
- The Use of ApplicationsVoice-driven editors and dictation apps are all examples of pipelines that can turn raw ASR data into text for humans to read.
What S1-mini does
The S1-mini text normalizer is not a transcription tool or a chat program. This is the next step after automatic speech recognition.
audio → ASR (Whisper, Parakeet, …) → S1-mini → clean text
Sayings are converted into written text by removing filler words and resolving false starts. It also capitalizes and punctuates spoken dates, currencies, time, and numbers. Say “support at superwhisper dot com” You get [email protected].
This model has been fine-tuned. Qwen/Qwen3-0.6B. The card has 28 layers, 8 key/value head with GQA and 596M unique parameter (0.44B without embedding), 16 query heads, and weights BF16. Hub reports 0.8B, but the embedded data is actually stored twice. The card clearly explains this discrepancy. Release v1 only covers English, with a recommended input of 1,000 tokens.
Control line refers to the whole interface
After the fixed prompt for the system, S1-mini displays a line of control, followed by an unprocessed transcript.
[Styling: ] [Structure: ] [Context: ]
Style Take a look at the following: You can also find out more about our services by clicking here., semi-casual, semi-formalIf you want to know more about. You can also find out more about the formal language of. The Structure of the Take a look at the following: prose The following are some examples of how to use You can find out more about this by clicking on the links below.. Context Take a look at the following: General The following are some examples of how to use Email. The three axes were trained independently. If you send values that are not in those set, or change the prompt wording, output may be garbled or degraded. Superwhisper exposes an additional five-stop slider in the app. “balanced” Open weights show four Styling values that are trained.
Design is another constraint. The model does not correct factual errors, add new content, soften swear words, or change dialect. Integrations must treat the empty string that is returned by filler-only input as valid.
There are two common settings that can break integration
First, enable_thinking=False It is necessary. Qwen3 uses the same chat template, but defaults to think on. Because S1 mini was taught to start the assistant with an empty screen, it is necessary that they use their thinking-off template. block. If you omit the flag, your output will be unusable.
Second, decode greedily. generation_config.json Ships Do_sample : Fal. Qwen3 has inherited Qwen3’s GGUF build. Temperature = 0.6, top_p = 0.95” top_k = 20 Pass temperature 0 on all requests. In llama.cpp, use --jinja The following are some examples of how to use --chat-template-kwargs '{"enable_thinking":false}' Instead of --reasoning-budget 0It is a degrading factor.
Reported evaluation
Superwhisper analyzed S1 mini on 7,519 recorded cases, spanning 104 transcripts. With a text editing error rate of 11,6%, the token accuracy measured on Q4_K_M is 94.8%. In email-formatted texts, it correctly recognizes the salutation line 99.3% the time and sign-off 97% of the times. It produces email addresses that are 92% accurate and matches the right output structure 97.6%. The model does not loop or truncate in less than 1%. It also correctly ignores the output when it isn’t needed. This is vendor reported numbers from an internal test, and not results by third parties.
There are two different cloud models
The hosted model of speech-to text is S1-Voice. Superwhisper can report transcriptions up to 46x quicker than spoken time. Most dictations of less than 30 seconds appear 0.32 seconds following your stop. It averages a 6.8% rate of word errors across eight datasets, including earnings calls and meeting audio. This drops to just 2.2% when using LibriSpeech. Superwhisper claims that the 6.8% was lowest among 15 models they tested. S1-Voice also scored higher than WisprFlow, scoring 83 points out of 100 for its combined metric.
S1-Language is the hosted instruction-following model for cleanup, formatting, and summarization, and it appears in the model picker alongside models from Anthropic, OpenAI, and Groq. Cohere Transcribe with S1 mini offline is the recommended default, as well as S1 voice plus S1 language in cloud.
Interactive explainer
You can switch the control lines and see how they affect output by using this embed. The model card contains every input/output combination.
function ctlLine(axis,val,fixed){
var o={Styling:”,Structure:”,Context:”};
o[axis]=val;
for(var key in fixed){o[key]=fixed[key];}
Return ‘[Styling: ‘+o.Styling+’] [Structure: ‘+o.Structure+’] [Context: ‘+o.Context+’]’;
}
function renderAxis(key){
var d=D[key],sel=d.opts[Math.min(2,d.opts.length-1)];
if(key===’structure’||key===’context’)sel=d.opts[0];
panel.innerHTML=”
+’
Line 1: Control line
‘
+”
+’
‘
+’
+’
Raw ASR transcription Line 2
‘+esc(d.input)+’
‘
+’
S1-mini output The tempor 0 is greedyFunction Apply(Key,Val)
‘
+’
‘;
var opts=document.getElementById(‘s1opts’);
d.opts.forEach(function(v){
var el=document.createElement(‘div’);
el.className=”s1-opt”+(v===sel?’ on’:”);
el.textContent=v;
el.onclick=function(){
[].forEach.call(opts.children,function(c){c.className=”s1-opt”;});
el.className=”s1-opt on”;
apply(key,v);
};
opts.appendChild(el);
});
note.textContent=d.note;
apply(key,sel);
}
function apply(key,val){
var d=D[key];
document.getElementById(‘s1cl’).innerHTML=ctlLine(d.axis,val,d.fixed);
type(document.getElementById(‘s1out’),d.out[val]);
}
function renderFix(){
Var h=”
FIX.forEach(function(r,i){
h+=’
‘+esc(r[0])+’
→
‘+esc(r[1])+’
‘;
});
h+=’
‘;
panel.innerHTML=h;
[].forEach.call(panel.querySelectorAll(‘.s1-row’),function(row){
row.onclick=function(){
row.className=”s1-row hit”;
setTimeout(function(){row.className=”s1-row”;},700);
};
});
note.textContent=”Measured on the BF16 weights with greedy decoding, under [Styling: semi-formal] [Structure: prose] [Context: general]. Filler-only input returns an empty string, and integrations should treat that as a valid result rather than a failure. The model resolves self-corrections to the value the speaker landed on, and renders spoken numbers, dates, currency and email addresses in written form.”;
Resize();
}
function esc(s){return s.replace(/&/g,’&’).replace(/,'<‘).replace(/>/g,’>’).replace(/n/g,’
‘);}
Function resize(){
try{
var h=document.getElementById(‘s1w’).offsetHeight+40;
parent.postMessage({s1height:h},’*’);
}catch(e){}
}
var tabs=document.querySelectorAll(‘.s1-tab’);
[].forEach.call(tabs,function(t){
t.onclick=function(){
[].forEach.call(tabs,function(x){x.className=”s1-tab”;});
t.className=”s1-tab on”;
var v=t.getAttribute(‘data-t’);
if(v===’fix’){renderFix();}else{renderAxis(v);}
Resize();
};
});
renderAxis(‘styling’);
setTimeout(resize,120);
setTimeout(resize,600);
window.addEventListener(‘resize’,resize);
})();
“>
The Key Takeaways
- The S1-mini text normalizer is not a chat or transcriber model, but a standard 0.6B weights open text for ASR.
- It is possible to deploy the Q4_K_M on devices with a CPU of a laptop.
- Only a system prompt with a control line on three axes is used for steering.
enable_thinking=FalseTemperature 0 and temperature 0 must be used; many integration problems can be traced to them.
Click here to find out more Model Weights The following are some examples of how to get started: Technical Details. Also, feel free to follow us on Twitter Don’t forget about our 150k+ML SubReddit Subscribe now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us


