Sign up

NANI-Nithin/Qwen3-TTS-12Hz-1.7B-Base-GGUF

🤗 Hugging Face sourcetext-to-speechapache-2.06.0 GBGGUF✓ HF checksums availableupdated today
No torrent yet →

Qwen3-TTS-12Hz-1.7B-Base — GGUF quantizations

Text-to-speech model quantized to llama.cpp GGUF, ready to run with llama-tts.

Source model: Qwen3-TTS-12Hz-1.7B-Base by the Qwen team — Qwen's 1.7B text-to-speech base model in the 12 Hz frame-rate variant. This repo contains only the quantized GGUF weights; all credit for the model itself belongs to the Qwen team. Consult the original Qwen model repository for licence and usage terms.

Files

File Size Notes
Qwen3-TTS-12Hz-1.7B-Base-Q4_K_M.gguf 0.96 GB The usual default. Best quality-per-byte for most people.
Qwen3-TTS-12Hz-1.7B-Base-Q5_K_M.gguf 1.14 GB Very good quality, noticeably smaller than Q6_K.
Qwen3-TTS-12Hz-1.7B-Base-Q6_K.gguf 1.33 GB Near-lossless; the last stop before quality becomes measurable.
Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf 1.72 GB Effectively lossless. Use when disk and RAM are not the constraint.
mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf 0.46 GB Required companion — audio projection model. Download alongside any main GGUF.

Total: ~5.62 GB.

Running

This is a voice bundle, not a single file: every llama-tts invocation needs the main GGUF and the mmproj companion file via -mm.

# via the Hub (all files in one command)
llama-tts -hf NANI-Nithin/Qwen3-TTS-12Hz-1.7B-Base-GGUF:Q4_K_M \
  -p "Hello, this is a test." --tts-lang en

# local files
llama-tts -m Qwen3-TTS-12Hz-1.7B-Base-Q4_K_M.gguf \
  -mm mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf \
  -p "Hello, this is a test." --tts-lang en

A speaker reference is optional. Output is 24 kHz mono WAV.

Validation

All quants were validated on release (2026-09-17) by actual llama-tts inference — 3 fixtures each: English numbers/punctuation, a long English passage, and a short French sentence:

  • Zero clipping on every fixture across all four quants
  • Long-passage round-trip ASR WER of 0.0 at Q5_K_M, Q6_K and Q8_0
  • Aggregate round-trip WER 0.085–0.217 depending on quant, dominated by ASR-side quirks on short multilingual fixtures, not audible artifacts
  • Every quant additionally passed a human listening review before publication

Measured on the release machine: ~1.9–2.0× real-time factor (~100–120 minutes of audio per wall-clock hour).

Which quant should I download?

  • Most users: Q4_K_M (0.96 GB) — smallest, reviewed and accepted.
  • Best quality: Q8_0, effectively lossless — pair it with the required mmproj companion file.
  • In between: Q5_K_M or Q6_K, both listener-accepted with no measurable quality complaints.

These files are llama.cpp GGUF and are not loaded by vLLM, SGLang or transformers the way the source model is.