Qwen/Qwen3-TTS Base 1.7B

• License: Apache-2.0
• Release date: 2026-01-22 (Qwen3-TTS series announcement)
• Modalities: Input: text, optional reference audio; Output: audio
• Original model: Qwen/Qwen3-TTS-12Hz-1.7B-Base

Description

Qwen3-TTS Base generates speech in ten languages and can imitate a speaker from a short reference recording. It uses a multi-codebook speech language model and the Qwen3-TTS 12 Hz tokenizer. This is the voice-cloning Base checkpoint, not the CustomVoice or VoiceDesign variant.

Run

self serve qwen3-tts:1.7b

Here is a link to the API docs: Audio API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{"input":"Welcome to the local speech model.","language":"en"}' \
  --output speech.wav

The response is saved as speech.wav (24 kHz WAV audio).

Engine Parameters

Served by the persistent ggmlc-audio engine through the ggmlc-audio adapter. These are startup settings, not per-request JSON fields; the registry sets no overrides, so the engine defaults apply.

Parameter	Default	What it controls
context_size	4096	Tokens shared by reference audio, text and generated frames; accepts 512 to 262144.
threads	Half the hardware concurrency (at least 1)	CPU workers; accepts 1 to 1024.
gpu_layers	-1	GPU-offloaded layers; -1 means all, 0 means CPU; accepts -1 to 10000. CPU device selection forces zero offload.
temperature	0.8	Token-sampler temperature; accepts 0.01 to 10.
top_p	0.95	Nucleus sampling; accepts 0 to 1.
top_k	40	Top-k sampling; accepts 1 to 1000 (1 selects greedy sampling).
frames	512	Maximum generated audio frames per request; accepts 1 to 100000.
seed	Random	RNG seed; accepts 0 to 4294967295. Reproducibility depends on the pipeline and device.

Override startup settings with --set <parameter>=<value>.

Quantizations and Files

Default variant: q4. Sizes and SHA-256 hashes are pinned in the model registry.

q4 (default)

File	Role	Size	SHA-256	Origin
Qwen3-TTS-12Hz-1.7B-Base-Q4_K_M.gguf	model	988.0 MiB (1035965280 bytes)	8d18c94acb2addd042f97da63c98be144eafa76d0d9495177eab65130cf85129	ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF
mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf	mmproj	425.7 MiB (446422912 bytes)	6fd65188839bcd6ecc91b277ad471e22a0edfada4699a0fe82f1165c18cfcce2	ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF

Total download: 1.38 GiB (1482388192 bytes).

q8

File	Role	Size	SHA-256	Origin
Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf	model	1.72 GiB (1847874400 bytes)	ac7931aeb2e7aad1a6ed6602d353a5679c9d096b18ce8204ac730a8408d572e1	ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF
mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf	mmproj	425.7 MiB (446422912 bytes)	6fd65188839bcd6ecc91b277ad471e22a0edfada4699a0fe82f1165c18cfcce2	ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF

Total download: 2.14 GiB (2294297312 bytes).

bf16

File	Role	Size	SHA-256	Origin
Qwen3-TTS-12Hz-1.7B-Base-bf16.gguf	model	3.23 GiB (3472593760 bytes)	0322c634ad5d3282524bc45bff030ab3f6f2a32ba14d5e7dace7eed75ecede46	ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF
mmproj-Qwen3-TTS-12Hz-1.7B-Base-bf16.gguf	mmproj	638.1 MiB (669081472 bytes)	b9503e95e44705739cf82c15ce909b040b96a31ad79d02cf4dc5e1484399265d	ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF

Total download: 3.86 GiB (4141675232 bytes).

Notes

• input is spoken content, not a chat or style prompt. Non-empty instructions, speed other than 1, non-WAV output and SSE streaming return 422 unsupported_capability. Emotion, pitch and loudness controls are not exposed.
• Without reference audio the runtime uses its default voice. OpenAI voice names such as alloy select that same voice, not different speakers. No reference voice file is bundled for this model.
• For cloning, provide a clean 5-20 second WAV or MP3 of one speaker as multipart ref_audio or base64 voice.audio. Send it on every request: no recording or reusable profile is stored. JSON filesystem paths and the upstream Python API's ref_text workflow are not supported here.
• The following example requires an existing local file named my-voice.mp3, containing a recording you own or have permission to use:

curl http://localhost:8080/v1/audio/speech \
  -F input='This sentence should sound like the reference speaker.' \
  -F language=en \
  -F ref_audio=@my-voice.mp3 \
  --output cloned.wav

• Set per-request language explicitly for English (en), German (de), French (fr), Spanish (es), Italian (it), Portuguese (pt), Russian (ru), Japanese (ja), Korean (ko) or Chinese (zh).
• The persistent engine keeps weights loaded between requests. Build from source with just engine audio if needed; supported builds include CPU, Vulkan, CUDA 12.8 and CUDA 13.4. Earlier local CPU measurements for q4 were under 1 second to become ready with cached files and about 4 seconds per short sentence; these are not performance guarantees.
• Both listed files are required for each variant. Download size is not the total runtime memory requirement. Origin links use main because these entries pin hashes, not repository revisions.

Quality benchmarks

Seed-TTS test set: Publisher-reported

Qwen/Qwen3-TTS-12Hz-1.7B-Base

Publisher-reported BF16, max_new_tokens=2048, default checkpoint sampling, language=auto. Not local Q4 results or an independent leaderboard rank.

Dataset	Metric	Score
Seed-TTS test-zh	WER (%)	0.77
Seed-TTS test-en	WER (%)	1.24

Benchmarks

No submitted benchmark reports yet.