aufklarer/Qwen3-TTS-12Hz-1.7B-Base-MLX-bf16

Qwen3 TTS 12Hz 1.7B Base — MLX bf16

Full-precision bf16 (non-quantized) MLX conversion of Qwen/Qwen3-TTS-12Hz-1.7B-Base for Apple Silicon inference — the highest-quality 1.7B variant.

Usage

Used by speech-swift Qwen3TTS module:

let model = try await Qwen3TTSModel.fromPretrained(
    modelId: "aufklarer/Qwen3-TTS-12Hz-1.7B-Base-MLX-bf16"
)
let audio = try model.synthesize("Hello, world!")

audio speak "Hello, world!" --model 1.7b -o output.wav

Model Details

• Architecture: Qwen3-TTS (Talker transformer + Code Predictor + speech tokenizer decoder)
• Parameters: 1.7B
• Precision: bf16 / fp16 — no quantization (plain Linear weights)
• Size: ~3.7 GB
• Sample rate: 24 kHz
• Codec rate: 12.5 Hz

Performance

Apple Silicon (M-series), MLX, 1.7B variants:

Precision	RTF	Peak RAM	Notes
8-bit	0.39	2.8 GiB	good
bf16	0.48	4.1 GiB	best quality

Round-trip WER (synthesize → Qwen3-ASR transcribe → WER), 15 sentences: 7.05 % (TTS + ASR round-trip; 0 synthesis failures).

Variants

Variant	Precision	Size	Model ID
0.6B 8-bit	8-bit	~1.3 GB	aufklarer/Qwen3-TTS-12Hz-0.6B-Base-MLX-8bit
1.7B 8-bit	8-bit	~2.8 GB	aufklarer/Qwen3-TTS-12Hz-1.7B-Base-MLX-8bit
1.7B bf16	bf16	~3.7 GB	aufklarer/Qwen3-TTS-12Hz-1.7B-Base-MLX-bf16

---

• Guide: soniqo.audio/guides/speak
• Docs: soniqo.audio
• GitHub: soniqo/speech-swift