CoolFace
Modelpublic

aufklarer/Voxtral-Mini-3B-2507-MLX-5bit

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
0likes40downloads
Model Card

Voxtral Mini 3B 2507 — MLX 5-bit affine

MLX weights for Voxtral Mini 3B 2507, ready for local speech-to-text with speech-swift.

The FP16, INT5, and INT8 family passed the isolated WER, RTF, throughput, RSS, and physical-footprint compatibility gate on 2026-07-22 using English FLEURS.

ParametersPrecisionSample rateLanguages
3B5-bit affine16 kHz8

Files

FileDescription
model*.safetensorsMLX model weights
config.json and params.jsonModel and runtime configuration
tekken.jsonTekken tokenizer data

Usage

bash
speech transcribe recording.wav --engine voxtral --model int5
swift
import VoxtralASR

let model = try await VoxtralModel.load(
    "aufklarer/Voxtral-Mini-3B-2507-MLX-5bit"
)
let text = model.transcribe(
    audio: samples,
    sampleRate: sourceSampleRate,
    language: "en"
)

Benchmark

English FLEURS read speech, 80 utterances (759.56 seconds), Apple M5 Pro:

VariantBundleWERMean RTFPeak RSS
FP168.71 GiB4.633%0.13059,091 MiB
INT53.77 GiB4.744%0.07394,035 MiB
INT85.18 GiB4.578%0.09065,487 MiB

These read-speech measurements are not Artificial Analysis Conversational or ElevenLabs Scribe results.

Source and license

The base model is mistralai/Voxtral-Mini-3B-2507 and is released under Apache 2.0.

Links