What it is
Qwen3-TTS is a family of open text-to-speech models released by Alibaba's Qwen team on January 22, 2026, in 0.6-billion and 1.7-billion-parameter sizes under Apache 2.0. It covers 10 languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. The weights were last updated on January 29, 2026, and Alibaba's hosted line has since moved to the separate Qwen-Audio-3.x TTS models.
There are three flavours. VoiceDesign builds a new voice from a written description, CustomVoice offers 9 preset voices whose tone can be steered with instructions, and Base clones a voice from 3 seconds of reference audio. Qwen says the streaming design can deliver first audio after a single character of input, with end-to-end latency as low as 97 ms.
Our review
Qwen3-TTS offers the broadest set of controls among open speech models, and the weakest protections. Voice design from a written prompt and instruction-driven tone are genuinely useful, and Apache 2.0 removes licensing friction. That combination explains its reach: the 1.7B Base cloning model alone drew about 2.9 million Hugging Face downloads in the past month, and the CustomVoice model about 2 million.
Independent listening tests are less kind. On Artificial Analysis's speech arena, Alibaba's hosted Qwen3 TTS ranks eighty-seventh of 98 at 930 Elo, and Qwen3 TTS Flash eighty-sixth at 943. Kokoro, an 82-million-parameter model, scores 1065. Those are the API versions, which may not match the open weights exactly, and Qwen's own word-error-rate tables are self-reported. Alibaba's newer Qwen-Audio-3.1-TTS-Plus sits third at 1301, a model Alibaba sells through its API rather than as Qwen3-TTS weights.
The Base models clone a voice from 3 seconds of audio plus its transcript, yet neither the model card nor the GitHub README mentions consent, misuse or watermarking. Chatterbox, by contrast, watermarks every clip. Anyone exposing Qwen3-TTS cloning to the public is taking on impersonation risk the model does nothing to reduce.
The models are small enough to be cheap to run: the 1.7B Base checkpoint is about 1.9 billion parameters, roughly 4GB in BF16, within reach of a single consumer GPU. The catch is serving. vLLM-Omni supports it only for offline inference so far, so a streaming production endpoint means writing your own server around the qwen-tts package.
Where it earns its place
A backend engineer at a European language-learning app
Generate example sentences in 10 languages with consistent, describable voices
Use VoiceDesign to define a voice per course ('warm, slow, mid-30s teacher') and batch-render offline on one GPU. Apache 2.0 means no per-character fees; listener rankings suggest a quality ceiling below hosted rivals.
A product manager at a 20-person audiobook start-up
Prototype narration cheaply before paying for a premium voice API
Render drafts on Model Studio's qwen3-tts-flash at $0.10 per 10,000 characters, so a 500,000-character novel costs about $5, then compare against hosted leaders that cost several times more per character.
A trust-and-safety lead at a voice-agent company
Decide whether to expose self-serve voice cloning
Qwen3-TTS Base clones from 3 seconds of audio and ships with no consent flow or watermark, so the company would need its own speaker verification, disclosure and logging before launch. Chatterbox's built-in watermark is the reference point.
Before you commit
- Setup
- pip install the qwen-tts package; weights download by model name. Qwen recommends FlashAttention, and its Gradio demo needs HTTPS for microphone input when cloning.
- Integrations
- qwen-tts Python package, Transformers, vLLM-Omni (offline only), Alibaba Model Studio API in Singapore and Beijing regions.
- Data and privacy
- Self-hosted under Apache 2.0, with no usage restrictions beyond the licence and nothing leaving your servers. On Model Studio, non-streaming results are returned as audio URLs valid for 24 hours.
- Cost at scale
- Self-hosted, the 1.7B model is about 4GB in BF16 and fits a single consumer GPU. Hosted, 100 million characters a month on qwen3-tts-flash costs about $1,000, or $1,300 on the realtime variant.
- Limits
- 10 languages; 9 preset voices on CustomVoice; no published watermark or misuse controls; online vLLM serving not yet supported; the hosted models rank in the bottom 15 of Artificial Analysis's arena.
What's good
- Apache 2.0 in 0.6B and 1.7B sizes, about 4GB in BF16 for the larger
- Voice design from a text description, plus 9 steerable preset voices
- Hosted Flash costs $0.10 per 10,000 characters
- Streaming with first audio after one character, per Qwen
What's not
- Hosted versions rank eighty-sixth and eighty-seventh of 98 on Artificial Analysis
- 3-second cloning with no consent checks or watermark
- vLLM-Omni supports offline inference only
- Weights unchanged since January 2026
Best for
Developers building multilingual voice features who want fine control under a permissive licence and will add their own safeguards.
Alternatives
Qwen3-TTS: common questions
- Is Qwen3-TTS worth it?
- Qwen3-TTS hands developers fine-grained voice controls and a three-second cloning tool with no guardrails, at a quality listeners rank near the bottom. We rate it 3.5 out of 5.
- How much does Qwen3-TTS cost?
- Free under Apache 2.0. Alibaba Model Studio (Singapore): qwen3-tts-flash $0.10 and qwen3-tts-flash-realtime $0.13 per 10,000 input characters, with a 90-day free quota.
- What are the downsides of Qwen3-TTS?
- Hosted versions rank eighty-sixth and eighty-seventh of 98 on Artificial Analysis; 3-second cloning with no consent checks or watermark; vLLM-Omni supports offline inference only; weights unchanged since January 2026.
- Who is Qwen3-TTS best for?
- Developers building multilingual voice features who want fine control under a permissive licence and will add their own safeguards.
- How much does Qwen3-TTS cost for a team or in production?
- Self-hosted, the 1.7B model is about 4GB in BF16 and fits a single consumer GPU. Hosted, 100 million characters a month on qwen3-tts-flash costs about $1,000, or $1,300 on the realtime variant.
- Is Qwen3-TTS safe for company data?
- Self-hosted under Apache 2.0, with no usage restrictions beyond the licence and nothing leaving your servers. On Model Studio, non-streaming results are returned as audio URLs valid for 24 hours.
- What are the best Qwen3-TTS alternatives?
- The closest alternatives we have reviewed are Chatterbox, Kokoro-82M and Gemini 3.8 Flash TTS.
Answers checked Oct 22, 2026.