Together AI announced that Qwen3 TTS, an open text-to-speech model from Qwen, can now be run on its platform, both on Together Serverless Inference and on dedicated endpoints. Serverless use costs $20 per 1M characters.
What the model does
The model keeps the identity of the speaker apart from the manner of speaking. You pick one of nine preset voices, and that voice stays recognisable while a short text instruction changes its delivery. The source gives examples such as "excited and upbeat", "angry and impatient" and "sad and disappointed", all applied to the same line read by the same voice.
Every preset speaks the same 10 languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. You supply the text already in the target language, since the model does not translate. Each voice has a native language, and Qwen advises using a voice in that language for the best quality. Of the nine, two are English, four are Chinese (two of them in Beijing and Sichuan dialects), one is Japanese and one is Korean.
Speed and streaming
Together says its inference team tuned the serving stack so that first audio can arrive in as little as 97ms. That is the vendor's own figure, measured when the response is streamed, and we have no independent test of it. Two streaming routes exist:
- HTTP streaming for cases where the full text is known up front. Audio comes back as server-sent events carrying encoded chunks of raw PCM.
- WebSocket for cases where text arrives piece by piece, for example from a language model generating a reply. You append text as it appears and commit when it is complete, and the server synthesises each sentence once it holds the whole sentence.
Practical limits
Output formats are wav, mp3, raw PCM and mulaw, at 24000 or 8000 sample rate. Opus, aac and flac are not supported, and word-level timestamps are not available for this model. The instruction field is named instructions; a misspelled field is skipped silently, so a voice that ignores your tone request is likely a typo.
For choosing a voice, Together points to its VoiceFinder tool, where you can select the Qwen3 TTS model, preview each preset and generate up to 500 characters of your own text. Together also states that Qwen released the 1.7B model under the Apache 2.0 licence, with weights on Hugging Face.
What is not in the announcement
The post gives no dedicated-endpoint pricing, no usage limits and no quality comparison with other speech models. Teams weighing it against a current voice provider will need to test with their own scripts and measure latency in their own setup.