Qwen3-TTS Serving Beats the Official Endpoint by 11×

Two entities serve the same open-weights Qwen3-TTS 1.7B model. Nari Labs' Fast endpoint: 63 ms median time-to-first-audio, 3.8% WER, $10/1M characters. Alibaba's official Qwen3 TTS Flash Realtime endpoint: 692 ms TTFA, 8.8% WER. Same weights. Eleven times the latency, 2.3× the error rate, and the official one isn't even cheaper per the Coval pricing directory. On the ASR side the same story: Nari's Qwen3-ASR Fast does 44 ms time-to-final-segment at 3.6% WER for $0.12/hour, while AssemblyAI's Universal 3.5 Pro — which only edges it on WER at 3.5% — costs 3.75× more.

What shipped

Nari Labs (the Dia TTS team) launched Qwen3-TTS and Qwen3-ASR, both 1.7B, both public beta, and posted Coval leaderboard numbers today: #1 WER and #2 TTFA on the TTS benchmark, #1 TTFS and #2 WER on STT — top of the quality-latency and cost Pareto frontiers among publicly available endpoints. Coval re-runs benchmarks every 30 minutes, so these are snapshots, but the ordering against ElevenLabs (5× the price) and Cartesia Sonic 3.6 (6.5×) has more headroom than the gap Coval's fluctuation would explain.

Why it matters

The interesting number isn't the ranking — it's the self-comparison Nari included, almost in passing. Official Qwen weights served by Alibaba: 692 ms, 8.8% WER. The same weights served properly: 63 ms, 3.8%. That's the open-weights thesis sharpened into a single data point: when the weights are public, the serving stack is the entire moat. Model providers competing with third parties serving their own open checkpoints are competing against their own artifacts. HN commenters poked at the demo quality — mid-clip voice switching, pacing that sounds like 1.25× playback — which is fair; WER doesn't catch prosody. But WER-at-latency is what production voice agents gate on, and on that axis the checkpoint you download beats the API from the lab that trained it.

Verdict

Watch this pattern, not this benchmark. Every time an open-weights audio model ships, expect third-party serving to out-pace the official endpoint within weeks. If Alibaba can't serve Qwen3-TTS better than Nari can, the "official API" premium on open models is officially dead.