QWEN / STREAMING TTS / 2026

Qwen3-TTS-12Hz-1.7B

TOTAL1.93Bparameters

ACTIVE1.93Bruntime path

DEPTH28 + 5repeated blocks

HIDDEN2,048main trunk

CONTEXT32Kstream / sequence

CHECKPOINT4.54GB2 shards · Apache-2.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

128 tokens → 10 seconds · 932.38G arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT128TOKENS

TRANSFORMER STACK28 + 5 LAYERSTalker generates semantic stream · predictor fills codebooks

Dual-Track Residual

OUTPUT10SECONDS

wraps every repeated block

CAUSAL GQA · DUAL-TRACK STREAMING文本输入与音频输出分别推进，使尚未读完的文本与已生成音频可并行。

Code Predictor在主语义 token 后补齐多码本声学细节。

TALKER28L2,048

PREDICT5L1,024

CODE RATE12.5Hzaudio

GROUPS16codebooks

LATENCY≈97msfirst packet

PARAMS1.93Bheader