Qwen3 TTS arrives on Together AI with first audio in 97 ms and 9 voices in 10 languages

Together AI announced on Thursday (20/11) that it now offers Qwen3 TTS, the open text-to-speech model from Alibaba's Qwen team, on its serverless inference platform — with first audio in just 97 ms, according to the company. The model, identified as Qwen/Qwen3-TTS-12Hz-1.7B, is distributed under the Apache 2.0 license, with weights on Hugging Face, and is also available on Together AI dedicated endpoints.

One voice, ten languages — and tone by instruction

Qwen3 TTS's key feature, according to Together AI, is separating who speaks from how they speak: the developer picks one of nine preset voices — aiden and ryan (English), vivian, serena, uncle_fu, dylan and eric (Chinese), ono_anna (Japanese) and sohee (Korean) — and that voice stays the same while the delivery changes through a plain-language instruction, such as “excited and upbeat”. Each of the nine voices speaks all 10 supported languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. Qwen recommends using each voice in its native language for the best quality.

Built for real-time voice agents

Latency is the core of the offer: Together AI's inference team tuned the serving stack so speech starts streaming fast enough for real-time conversation — first audio arrives in as little as 97 ms when the response is streamed. The model supports streaming in two ways: over HTTP, with server-sent events as the audio is generated, and over WebSocket, for when the text is still being written token by token by a language model. Output formats are wav, mp3, raw (PCM) or mulaw, at a 24 kHz sample rate by default. To compare voices before choosing, Together AI offers the VoiceFinder tool, which plays samples and generates up to 500 characters of your own text in each voice.

Price and context

On Together AI serverless inference, Qwen3 TTS costs $20 per 1 million characters. One clarification: the 20 November news is the hosted offering — the model itself was already circulating as Qwen open weights. It is another chapter for the Qwen family in 2026: Geekiki covered this week Qwen3.5-27B improving by 6.6 points by rewriting its own weights.

Image credit: Qwen logo, Alibaba Cloud, via Wikimedia Commons, Apache 2.0 license.

Sources: Together AI (blog oficial); OpenRouter; Futurum Group

Transparency: this article was produced with the assistance of artificial intelligence, based on cross-checking public posts on X and internet sources, and reviewed editorially before publication. Please consult the original sources to verify the facts.

By Muse