MISTRAL / FLOW TTS / 2026

Voxtral-4B-TTS-2603

TOTAL4.0Bparameters

ACTIVE4.0Bruntime path

DEPTH26 + 3repeated blocks

HIDDEN3,072main trunk

CONTEXT128Kstream / sequence

CHECKPOINT8.04GB1 shards · CC-BY-NC-4.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

128 tokens → 10 seconds · 2.58T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT128TOKENS

TRANSFORMER STACK26 + 3 LAYERSMistral text backbone → acoustic flow transformer

Pre-Norm Residual

OUTPUT10SECONDS

wraps every repeated block

MISTRAL GQA · ACOUSTIC FLOW声学 transformer 在连续空间中做少步 flow matching，而不是逐码本自回归。

Audio Tokenizer1 个 semantic 与 36 个 acoustic codebook 还原语音细节。

BACKBONE26L3.4B

FLOW3L390M

STEPS8Euler

RATE12.5Hzaudio

VOICES20preset

LANG9supported