What was released
Mistral has published Voxtral 3, a family of speech models covering transcription, streaming transcription and synthesis, on its blog. The line-up is Voxtral Small Transcribe 3 for batch accuracy, Voxtral Mini Transcribe 3 at 4 billion parameters for efficiency, Voxtral Mini Transcribe Realtime 3 for streaming, and Voxtral Speech 3 for text-to-speech. A voice agent SDK built on LiveKit chains them into a single pipeline.
Batch transcription covers more than 50 languages, including the 24 official EU languages; the realtime model covers more than 35; and the speech model covers more than 35 languages and 50 dialects, with voice presets in 18. All the numbers below are Mistral’s own, published on its product page rather than measured independently.
The latency tax, in Mistral’s own figures
The most useful thing on the page is the realtime chart, because it prices speed rather than claiming it away. Measured on FLEURS across 35 languages, the Mini Transcribe 3 offline baseline posts a macro-average word error rate of 4.0%. Running the same model as a stream, the error climbs with every step down in latency: 5.7% at 2,400ms, 6.0% at 960ms, 6.4% at 480ms and 7.4% at 240ms.
That is a direct trade published by the vendor: a quarter-second response costs roughly 85% more word errors than the same model given the whole file. Mistral places Gemini 3.5 Realtime Transcribe at 4.2% on the same chart, Microsoft AI Transcribe 2 Streaming at 5.8%, GPT Live Transcribe at 8.4% and Scribe v2 Realtime at 8.7%.
On batch transcription, Mistral reports Voxtral Small Transcribe 3 at about 4.1% word error with the language forced and 4.0% with it auto-detected, against about 16.0% and 14.9% for Whisper large-v3 — the oldest comparison on the chart, and the one that flatters the result most.
Where it says it beats the field
The diarisation numbers are the stronger claim. Averaged over five conversational sets, Mistral reports a concatenated-minimum word error rate of 18.4% for Voxtral Small Transcribe 3 and 18.6% for the Mini, against 19.0% for Pyannote’s Precision 3 orchestrator, 20.3% for Microsoft AI Transcribe 2, 23.0% for Scribe v2 and 38.4% for Gemini 3.5 Transcribe. The previous generation, Voxtral Mini Transcribe 2, sits at 30.9% — the largest single improvement on the page, and one that is Mistral measuring itself against itself.
Prices and plumbing
Mistral is charging $0.0035 a minute for Mini Transcribe 3, $0.0040 for Small Transcribe 3 and $0.008 for the realtime model, with speech synthesis at $13 per million characters. Voice cloning needs roughly 30 minutes of sampled audio. The transcription models support diarisation, keyterm biasing and end-of-turn detection; the agent stack reaches phones over SIP and browsers over WebRTC, and Mistral says a full agent turn completes in under a second.
The models are available through Mistral’s API on European, US and global endpoints, in the Studio playgrounds, and for self-deployment on request.
What is missing
For a company whose identity is built on open weights, the page is silent on licensing: it does not say whether any Voxtral 3 weights are released, under what terms, or whether the family is API-only. It also gives no maximum audio length, no context limit, no parameter counts for the Small transcription model or the speech model, and no explicit release date — the only date on the page is a footnote saying the speech evaluations were generated on 20 October.
The figures to wait for are the independent ones. Mistral’s realtime chart is unusually honest about what streaming costs; whether the batch and diarisation leads survive someone else’s test set is the part that is not yet on the record.
