Voxtral Mini 3B 2507
此 Soniqo 页面记录本地 speech-swift / speech-core 实现中的 Voxtral Mini 3B 2507。Hugging Face 包链接放在集成说明之后。
先进入站内页面
首页卡片和文档菜单先指向这里；源模型和权重包链接仍在本页提供。
概览
| 模型 | Voxtral Mini 3B 2507 |
|---|---|
| 用途 | High-accuracy multilingual offline speech-to-text |
| 后端 | Native MLX on Apple Silicon |
| 输出 | Plain-text transcription |
| 语言 | English, French, German, Spanish, Italian, Portuguese, Dutch, and Hindi |
| 许可证 | Apache-2.0 |
| 状态 | Published FP16, INT5, and INT8 bundles; INT5 is the default |
| 来源 | Mistral Voxtral Mini 3B 2507 |
| Swift 产品 | VoxtralASR |
| CLI / 运行时 | speech transcribe --engine voxtral |
使用
下面的片段对应当前 speech-swift 仓库暴露的 API 或命令。
# 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基准测试
测量于 2026-07-22，使用 80 条英语 FLEURS 朗读语音（共 759.56 秒），在配备 48 GB 内存、运行 macOS 26.5.2 的 Apple M5 Pro 上完成。每个变体都以 release 构建在独立进程中运行。
| 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 |
物理占用是与部署相关的统一内存指标，因为 MLX 可能会内存映射权重文件。FLEURS 为英语朗读语音：这些数字不代表在对话、电话、嘈杂或多说话人音频上的同等表现。
模型链接
实现说明
- 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.