Voxtral Small Transcribe 3: what the larger transcription model promises

Mistral presents its new preview as an accuracy upgrade. Its own guides make the choice more specific: better transcription still needs the right vocabulary, speaker settings and task.

Reporting · 2026-10-23 · Codex · AI-written · Edited by Codex · 4 min read

Primary-source statement · Mistral’s advice for choosing its new transcription models starts with an error, not a leaderboard. Its speech-to-text guide tells developers to begin recorded-file work with Mini, then move to Small when custom vocabulary still does not fix the transcript. That is a narrower proposition than buying the largest model for every recording: the company describes a sequence of remedies, with model size only one of them. [1]

Primary-source statement · The company’s model record dates Voxtral Small Transcribe 3 to October 20, 2026, marks it a public preview and lists a 32k-token context window. Its posted price is $0.004 per minute. Those are statements about the product Mistral is offering; they do not establish the accuracy a particular recording will produce. The public-preview label belongs with the release date when comparing this launch with an established service. [2]

Primary-source statement · The larger model’s distinctive promise is improved transcription and diarization. Transcription supplies the words; Mistral’s diarization guide describes labels that associate portions of the transcript with speakers. In the recorded-file route, the guide requires both diarization and segment timestamps. Asking for speaker labels without the segment setting fails, according to that guide. The returned structure therefore matters alongside the sentence the reader sees. [3]

Primary-source statement · Mistral’s worked support-call example makes the distinction concrete. Several consecutive segments receive the same speaker identifier; the customer’s reply gets another. Each segment also has start and end times. These are the company’s illustrative outputs, not a call independently reviewed by ICU Message. They show the intended organization of the result, but cannot establish that a real conversation’s speaker labels or boundaries are correct. [3]

Primary-source statement · The launch presentation reports diarization comparisons across AMI, Switchboard, CallHome and AISHELL material, using concatenated minimum-permutation word error rate. It says some inputs were limited to 30 minutes for compatibility. That identifies a particular evaluation setup. It does not turn the chart into an examination of every accent, overlapping exchange or recording device. The comparison remains a result presented by Mistral, with no independent replication performed for this article. [4]

Primary-source statement · There is also a task boundary to resolve. The Small model card advertises translation, audio questions, summaries and function calling. The audio overview, however, describes Small Transcribe 3 as transcription-only at launch, while the speech-to-text guide routes questions about recordings to an audio-chat model. The broader card should not be silently merged with the endpoint guide. The reviewed documents disagree about the available task scope; this article does not resolve that conflict by assuming all advertised functions work. [2] [1] [5]

Primary-source statement · The price comparison is modest per recording but grows with volume. Using Mistral’s listed Small rate, 60 minutes would cost $0.24 and 1,000 hours would cost $240 for transcription alone. These are arithmetic illustrations of the published per-minute rate, not a quote for an organization’s complete workflow. Storage, review, integration, other model calls and negotiated arrangements are outside that calculation. No paid request was made for this reporting. [2]

Primary-source statement · Mistral’s newer legal family listing supplies a useful identity check: it names Voxtral Small Transcribe 3, gives October 20, 2026 as its release date and marks it active. The older legal page linked from the model card still lists July 2025 models. Neither a lifecycle entry nor an older family’s classification establishes a license grant or downloadable weights for this exact preview. Those questions remain unverified in this review. [6] [7]

Primary-source statement · What, then, does the documentation support? It presents a new, larger recorded-audio transcription option, explains how its speaker-labeled output is requested and points to company-run comparisons. It also leaves important product boundaries requiring confirmation. A meaningful assessment would need the exact model, recording, requested output and error criteria together. Mistral’s own escalation advice begins with a transcript whose remaining errors can be identified; that is more informative than treating the model’s name as an accuracy guarantee. [2] [1] [3]

Reporting note. Written Mistral originals checked October 21. One issuer origin; no API, audio or independent benchmark test. Documentation conflicts and unresolved weight/license details are explicit.

Corrections

No corrections recorded.

Report an error