Alibaba's Qwen TTS tops AI leaderboard at $27.59 per 1M chars — but it's slow

Alibaba’s Tongyi Lab just dropped Qwen-Audio-3.0-TTS, and it’s already snagged the top spot on the Artificial Analysis quality leaderboard. That alone is turning heads, but the real story is how they got there: it’s not just another voice clone. They shipped two variants. Flash is built for real-time interaction, with first-packet latency hovering around 300 milliseconds. Plus prioritizes quality, and it’s the one that’s now in a statistical dead heat for first place with an Elo near 1,236, just nosing ahead of Simba 3.2. Both models are hosted through Alibaba Cloud Model Studio, so don’t expect to download any weights.

What’s genuinely new here is the control scheme. The team baked 86 fine-grained inline tags directly into the text stream. You can drop markers like [excited], [whispers], or even [laughing] and [clears throat] right into a sentence to guide delivery. It’s a thoughtful touch for developers who’ve been stitching together audio files to get that level of nuance. The model also handles 16 languages and 20 Chinese dialect regions, with the Flash variant posting the lowest average word error rate—a clean 3.87 across the board. A curated preset voice library ships with it, so you can get started without needing a pristine reference clip.

The price is aggressive. At $27.59 per million characters, it’s roughly a third of what ElevenLabs and MiniMax charge for tiers it now outranks on quality. That’s a direct shot across the bow for the incumbent API providers. But the trade-off is speed. Plus generates around 16 characters per second. That’s not a typo. Compare that to Sonic 3.5 at 120 characters per second, and you see the bottleneck immediately. It’s a stat that will matter a lot if you’re generating long-form narration or need high throughput.

Early developer chatter is cautiously enthusiastic. The most common gripes are predictable: it’s a hosted-only model, and that throughput figure is a real limitation for certain workloads. But the signal is clear—a non-Western lab just set the quality benchmark while undercutting on cost. The model is accessible via a bidirectional WebSocket streaming protocol, with SDKs for Python, Java, Go, and more. If Alibaba can boost the generation speed without sacrificing the quality that got it to the top of the leaderboard, this becomes a much bigger problem for the incumbents.