Qwen3 TTS 1.7B goes live on Together AI with instruction-controlled tone, 97ms first audio
togethercompute · x · 2026-11-23
Together AI is hosting Alibaba's Apache 2.0-licensed Qwen3 TTS 12Hz 1.7B model at $20 per 1M characters.
- Covers 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian) plus dialect voice profiles, with 9 preset voices that can each speak any supported language
- Key feature is instruction control: keeping the text identical and only changing the instructions field in the request makes the same voice deliver different tone, pace, and emotion
- Built on Qwen's 12Hz speech tokenizer with a single end-to-end LM architecture, supports streaming with first audio in as little as 97ms, and holds up on noisy input text
Related event: Together AI Launches Qwen3 TTS with Instruction-Controlled Tone(2 posts)→
More from Multimodal
- With just two reference images, Fable conjures a full retro music player UI with scrolling lyrics — Aryvyo · 2026-11-23
- A 3D Christmas village built with Three.js, Blender and Meshy AI introduces Milo and Achoo — techartist_ · 2026-11-23
- AI-generated 'The Emerald Helm' short stuns Reddit with cinematic knight visuals — PsychologicalSet4573 · 2026-11-23
- Claude Code + video MCP: 4.5 hours of GTA 6 footage turned into 8 clips in seconds — aryanXmahajan · 2026-11-23
- Dictating a full song and music video into existence via Claude Duet and Suno MCP — TJLarkin23 · 2026-11-23
- Qwen3 TTS: change how a voice delivers a line via the instructions field, no voice swap needed — togethercompute · 2026-11-23