Alibaba Cloud's Qwen team released Qwen3-TTS 0.6B on January 22, an open-source text-to-speech model designed for edge and mobile deployment. The model delivers 97ms latency at 12Hz streaming, supports 10 languages, and clones voices from 3-second audio samples.
The 0.6B variant weighs 2.52GB, roughly half the size of its 1.7B sibling (4.54GB). The team claims cross-language voice cloning: a Chinese voice sample can generate English, Japanese, or Korean speech. Nine premium voices ship with the model, offering control over gender, age, and dialect via text prompts.
The model competes directly with commercial services. Azure Speech Services and ElevenLabs charge per character or require cloud connectivity. Qwen3-TTS runs locally under Apache 2.0 license, trained on 5 million hours of speech data using what Alibaba calls a "dual-track LM architecture."
Deployment targets include NVIDIA Jetson edge devices and ComfyUI integration for workflow automation. The GitHub repository shows active deployment on Ollama for local inference. Early adopters report memory footprint challenges on resource-constrained hardware, though the 0.6B size aims to address this.
Security researcher Simon Willison flagged voice cloning risks: browser-based demos make it trivial to clone voices without consent. The model's ease of access amplifies existing deepfake concerns, particularly for voice authentication systems.
What this means in practice: CTOs evaluating TTS infrastructure now have a credible open-source alternative to cloud services. The trade-off is deployment complexity versus API simplicity. For regulated industries requiring on-premises processing (healthcare, government), the Apache 2.0 license and local inference matter.
The real test comes when enterprises attempt production deployment. The 97ms latency claim needs validation under load. Memory optimization for edge devices remains an open question. Voice cloning quality versus commercial alternatives will determine adoption.
Alibaba's timing is notable. The TTS market is projected to exceed $10 billion by 2030. Open-sourcing a competitive model undercuts cloud pricing while building ecosystem lock-in around Qwen infrastructure. History suggests this works when the model is genuinely good enough.