Mistral AI just released Voxtral 3. Mistral Voxtral 3 is a speech model family paired with a voice agent SDK. It ships 3 transcription models, 1 text-to-speech model and an SDK built on LiveKit. One place for listening, reasoning, tool calls and speaking. All 4 models are in public preview on the Mistral API.
TL;DR
- Size: Voxtral Mini Transcribe 3 has 4B parameters. Small Transcribe 3 size: not disclosed. Both batch models list a 32k context.
- Runs on: Mistral API (EU, US and global endpoints) or your own infrastructure via Mistral sales. GPU and VRAM needs: not disclosed. Mistral says Mini suits resource-constrained environments.
- Performance: Top 3 on independent speech-to-text rankings, strongest on diarization and multilingual batch transcription.
- Best: 2.1% AA-WER for Small Transcribe 3, #3 on Artificial Analysis.
- Worst: Realtime at 240 ms: 7.4% WER on FLEURS 35, vs 4.2% for Gemini 3.5 Realtime.
What is in the Voxtral 3 family?
Mistral lists 4 models and 1 SDK:
All transcription models support diarization and context biasing. Context biasing takes key terms or free text to improve recall. Batch transcription covers 50 languages, including 24 official EU languages.
How does the Voxtral voice agent SDK work?
The SDK is built on LiveKit and runs 3 stages as 1 streaming pipeline. Realtime Transcribe 3 listens and detects when the caller stops speaking. A general LLM, such as Mistral Small 4, reasons and calls tools. Voxtral Speech 3 replies in the tone of the conversation so far.
Each turn completes in under 1 second. The agent stops talking when the caller cuts in. It handles phone calls over SIP and browser or app sessions over WebRTC. Because each stage is a separate model, teams can swap the LLM, voice or language.
‘}).join(”);setTimeout(function(){el.querySelectorAll(‘.fl’).forEach(function(f){f.style.width=f.getAttribute(‘data-w’)+’%’})},40)}
function draw(k){cur=k;rows(bars,D[k].rows,D[k].max);cap.textContent=D[k].cap;setTimeout(size,60)}
R.querySelectorAll(‘[data-m]’).forEach(function(b){b.onclick=function(){R.querySelectorAll(‘[data-m]’).forEach(function(x){x.classList.toggle(‘on’,x===b)});draw(b.getAttribute(‘data-m’))}});
/* realtime */
var V={240:7.4,480:6.4,2400:5.7},lat=”480″,rtEl=R.querySelector(‘#vxRt’);
function rt(l){lat=l;var list=[[‘Gemini 3.5 Realtime Transcribe’,4.2],[‘Voxtral Mini Realtime 3 @ ‘+(l===’2400′?’2,400′:l)+’ ms’,V[l],1],[‘MAI-Transcribe-2 Streaming’,5.8],[‘GPT Live Transcribe’,8.4],[‘Scribe v2 Realtime’,8.7]];list.sort(function(a,b){return a[1]-b[1]});rows(rtEl,list,10);setTimeout(size,60)}
R.querySelectorAll(‘[data-l]’).forEach(function(b){b.onclick=function(){R.querySelectorAll(‘[data-l]’).forEach(function(x){x.classList.toggle(‘on’,x===b)});rt(b.getAttribute(‘data-l’))}});
/* pricing */
var rg=R.querySelector(‘#vxRange’),mn=R.querySelector(‘#vxMin’);
function price(){var m=+rg.value;mn.textContent=m.toLocaleString(‘en-US’);R.querySelectorAll(‘.p[data-r]’).forEach(function(p){var v=m*parseFloat(p.getAttribute(‘data-r’));p.textContent=”$”+(v<100?v.toFixed(2).replace(/\.00$/,”):Math.round(v).toLocaleString(‘en-US’))})}
rg.addEventListener(‘input’,price);price();
draw(‘fleurs’);rt(‘480’);
window.addEventListener(‘load’,size);window.addEventListener(‘resize’,size);setTimeout(size,300);
})();
