Voxtral Mini 3B 2507

This first-party Soniqo page documents Voxtral Mini 3B 2507 from the local speech-swift / speech-core implementation. Hugging Face bundles are linked below after the integration notes.

Internal Page First

Landing cards and docs menus now point here first; source model and bundle links remain available from this page.

At a Glance

Model	Voxtral Mini 3B 2507
Role	High-accuracy multilingual offline speech-to-text
Backend	Native MLX on Apple Silicon
Output	Plain-text transcription
Languages	English, French, German, Spanish, Italian, Portuguese, Dutch, and Hindi
License	Apache-2.0
Status	Published FP16, INT5, and INT8 bundles; INT5 is the default
Source	Mistral Voxtral Mini 3B 2507
Swift product	`VoxtralASR`
CLI / runtime	`speech transcribe --engine voxtral`

Use

The snippet below mirrors the current speech-swift API or command exposed by the repo.

# INT5 is the default.
speech transcribe recording.wav --engine voxtral

# Select another published precision and pass a language hint.
speech transcribe recording.wav --engine voxtral --model int8 --language fr

# --model also accepts a Hugging Face model ID or a local directory.
speech transcribe recording.wav --engine voxtral --model /models/voxtral/int5

Benchmark

Measured on 2026-07-22 over 80 English FLEURS read-speech utterances totalling 759.56 seconds, on an Apple M5 Pro with 48 GB of memory running macOS 26.5.2. Each variant ran in its own process from a release build.

Variant	Bundle	WER	Δ WER	Mean RTF	Overall ×RT	Footprint
FP16	8.71 GiB	4.633%	—	0.1305	8.05×	10,568 MiB
INT5	3.77 GiB	4.744%	+0.110 pp	0.0739	14.47×	6,012 MiB
INT8	5.18 GiB	4.578%	-0.055 pp	0.0906	11.79×	7,233 MiB

Physical footprint is the deployment-relevant unified-memory figure, because MLX may memory-map weight files. FLEURS is English read speech: these numbers do not claim parity on conversational, telephone, noisy, or speaker-heavy audio.

Model Links

Implementation Notes

The audio frontend resamples mono Float32 PCM to 16 kHz and packs up to 30 seconds of audio per request.
INT5 is the default: on the validated English FLEURS run it used a 3.77 GiB bundle, 6,012 MiB physical footprint, and 0.0739 mean RTF.
The decoder projects only the final prompt state through the 131,072-token language-model head, reducing quantized-model RTF without changing transcripts.
MLX has no INT7 affine kernel; use INT8 as the supported higher-quality option.
Audio is padded to whole 30-second chunks and concatenated into one prompt, so prompt and decoder-cache memory grow with the number of chunks. Segment very long recordings until a chunk-merge path exists.
This is a non-streaming engine.