Qwen3-ForcedAligner-0.6B — MLX 8-bit
8-bit quantized version of Qwen/Qwen3-ForcedAligner-0.6B for Apple Silicon inference via MLX.
Predicts word-level timestamps for audio+text pairs in a single non-autoregressive forward pass.
Model Details
| Detail | Value |
|---|---|
| Audio encoder | 24 layers, 1024 dim, 16 heads, float16 |
| Text decoder | 28 layers, 1024 hidden, 16Q/8KV heads, 8-bit quantized (group_size=64) |
| Classify head | Linear(1024, 5000), float16 |
| Timestamp resolution | 80ms per class (5000 classes = 400s max) |
| Total size | ~1.4 GB |
Usage
let aligner = try await Qwen3ForcedAligner.fromPretrained(
modelId: "aufklarer/Qwen3-ForcedAligner-0.6B-8bit"
)
let aligned = aligner.align(
audio: samples, text: "Hello world", sampleRate: 24000
)
Variants
| Variant | Size | Model ID |
|---|---|---|
| 4-bit | ~979 MB | aufklarer/Qwen3-ForcedAligner-0.6B-4bit |
| 8-bit | ~1.4 GB | aufklarer/Qwen3-ForcedAligner-0.6B-8bit |
| bf16 | ~1.8 GB | aufklarer/Qwen3-ForcedAligner-0.6B-bf16 |
| CoreML INT4 | ~630 MB | aufklarer/Qwen3-ForcedAligner-0.6B-CoreML-INT4 |
| CoreML INT8 | ~1.0 GB | aufklarer/Qwen3-ForcedAligner-0.6B-CoreML-INT8 |
Links
- Swift library: soniqo/speech-swift
- Base model: Qwen/Qwen3-ForcedAligner-0.6B
- Guide: soniqo.audio/guides/align
- Docs: soniqo.audio
- GitHub: soniqo/speech-swift