In this video I tested Alibaba’s newly open-sourced Qwen3-TTS family and, honestly, it’s basically the best free text-to-speech model you can go for right now if you want to run everything locally. I break down the two model sizes (1.7B and 0.6B), how lightweight they are (around a few GB), and why the performance feels well-optimized even on a normal setup. I also go over the language support—10 mainstream languages like Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian—and I call out what’s missing (Arabic would’ve been nice, and maybe Hindi too). I then focus on what actually matters: the voice quality and control. I show examples with energy, emotion (laughs, sighs, crying), gradual intensity changes, and style prompting—because this is the kind of range you need for AI shorts, films, and dubbing that doesn’t sound like YouTube’s horrendous auto-dubbing. I also demo the Hugging Face Space for quick testing and then show my ComfyUI workflow, including rapid voice cloning from short audio clips (I tried stuff like Trump, Putin, and Optimus Prime). My takeaway: the results can be insanely strong, but you’ll need precise prompting and a few attempts to “hammer down” the exact delivery—and some things (like making a convincing old man voice or mixing design-voice control with cloned voices) still feel limited.