AI Tools
Qwen3-TTS vs ElevenLabs: Which Text-to-Speech Should You Pay For?
Qwen3-TTS vs ElevenLabs comes down to cost: Qwen3-TTS self-hosts at roughly $2 per 1M characters on a single H100, about 50x cheaper than the ElevenLabs API, per Nari Labs' August 2026 benchmark. That makes Alibaba's Apache-2.0 model the strongest open-source ElevenLabs alternative for teams with GPU capacity and engineering time.
What matters
- Qwen3-TTS 1.7B CustomVoice self-hosts at about $2 per 1M characters on a single H100 at full utilization, versus $100 per 1M for ElevenLabs V3 (Nari Labs, Aug 2026).
- Nari Labs' tuned serving reached sub-50 ms p95 time-to-first-audio at 10 requests per second, the only one of five engines tested to do so.
- Qwen3-TTS is Apache-2.0 licensed, supports 10 languages, and the Base variant clones a voice from a 3-second audio clip.
- ElevenLabs starts at $6/mo (Starter, 30k credits) with Pro at $99/mo; Cartesia starts at $5/mo (Pro, 100k credits).
- The ~$2 per 1M figure assumes full GPU utilization; idle capacity, networking, and ops overhead push real-world cost higher.
What is Qwen3-TTS?
Qwen3-TTS is an open-source text-to-speech model family from the Qwen team at Alibaba Cloud, first released in January 2026 under the Apache-2.0 license. The flagship 1.7B-parameter CustomVoice model supports 10 languages and 9 premium timbres, while a separate Base variant can clone a voice from a 3-second audio clip. It uses a discrete multi-codebook architecture with a 12Hz tokenizer for streaming generation; the official README cites end-to-end synthesis latency as low as 97 ms.
Adoption is measurable, not theoretical: the 1.7B CustomVoice checkpoint on Hugging Face has 2.27 million downloads and 1,899 likes, and the QwenLM/Qwen3-TTS GitHub repository has 13,063 stars. The Apache-2.0 license, confirmed in both the model card and the repository LICENSE file, permits commercial use, modification, and redistribution, including voice cloning and fine-tuning.
The benchmark that changed the cost math
Nari Labs, an inference-optimization shop, published an August 19, 2026 benchmark showing a tuned Qwen3-TTS 1.7B CustomVoice deployment holding sub-50 ms p95 time-to-first-audio (TTFA) at 10 requests per second on a single NVIDIA H100 SXM. Five engines were compared under Poisson open-loop traffic: Nari's, vLLM-Omni, SGLang-Omni, VoxServe, and M*. Nari's was the only one to stay under 50 ms; it keeps sub-50 ms through 10 RPS and stays below 100 ms even at 20 RPS, producing about 630 characters per second.
Cost is where the gap widens. At $4.29 per hour for a 1x H100 SXM instance, Nari estimates about $2 per 1M characters at full utilization, versus $100 per 1M for ElevenLabs V3 and $49 per 1M for Cartesia Sonic 3.5. The implementation and benchmark are open source on GitHub under Apache-2.0.
Qwen3-TTS vs ElevenLabs: pricing compared
ElevenLabs sells credits, not characters. The Free tier includes 10k credits per month, Starter costs $6/mo for 30k, Creator is $22/mo (first month $11) for 121k, Pro is $99/mo for 600k, and Scale and Business climb to 1.8M and 6M credits. Cartesia's tiers are $0, $5, $49, and $299 per month for 20k, 100k, 1.25M, and 8M credits, with the current model listed as Sonic-3.6. Neither vendor publishes a per-character rate on its pricing page, so the per-1M figures below come from the Nari Labs benchmark, which used API-character costs.
For light usage, the API subscription math wins outright: a $6/mo ElevenLabs Starter plan costs less than renting any GPU for a month. Self-hosting only pays off when a single H100 is kept busy close to 24/7.
Who should self-host Qwen3-TTS
Self-hosting makes sense when you have a stable GPU workload. A streaming TTS service that runs near 24/7 can amortize the roughly $4.29/hr H100 cost across enough characters to beat API pricing by an order of magnitude. Teams already running vLLM or Kubernetes GPU fleets, or building voice agents with high call volumes, are the natural buyers.
Stay on ElevenLabs or Cartesia if you have no GPU capacity, want instant voice cloning with zero ops, need 192 kbps output or 44.1 kHz PCM (ElevenLabs Pro features), or generate only a few hours of audio per month. The free tiers of both APIs cover light usage.
Honest limits of the benchmark
The ~$2 per 1M figure is a full-utilization estimate; Nari's own footnote excludes networking, idle capacity, and operational overhead. A lightly loaded GPU costs the same per hour as a busy one. The benchmark also measures latency and throughput, not voice quality: no human evaluation is included, and output at sub-50 ms may differ from the reference implementation.
Two more caveats. The current Cartesia flagship is Sonic-3.6, a generation newer than the Sonic 3.5 the benchmark priced. And neither ElevenLabs nor Cartesia publishes TTFA figures, so the latency gap rests on one-sided data.
At a glance
| Feature | Qwen3-TTS 1.7B (self-hosted) | ElevenLabs (API) | Cartesia Sonic (API) |
|---|---|---|---|
| License | Apache-2.0 (verified on model card + LICENSE) | Proprietary | Proprietary |
| Est. cost per 1M chars | ~$2 at full H100 utilization (Nari est.) | ~$100 for V3 (Nari est.) | ~$49 for Sonic 3.5 (Nari est.) |
| Entry plan | Free weights + GPU time | $6/mo Starter, 30k credits | $5/mo Pro, 100k credits |
| Streaming TTFA | Sub-50 ms p95 at 10 RPS (Nari tuned) | Not published | Not published |
| Voice cloning | 3-second clone (Base model) | Instant Voice Cloning | Instant voice cloning |
| Languages | 10 | Not stated on pricing page | Not stated on pricing page |
FAQ
Is Qwen3-TTS free to use commercially?
Yes. Qwen3-TTS is released under Apache-2.0, confirmed in the Hugging Face model card and the GitHub LICENSE file, which permits commercial use, modification, and redistribution. The voice clone Base models carry the same license.
How much does Qwen3-TTS cost to run?
The weights are free; running them costs GPU time. Nari Labs estimates about $2 per 1M characters on a single H100 at full utilization (a $4.29/hr instance), excluding networking and idle overhead. ElevenLabs V3 runs about $100 per 1M characters on the API.
Can Qwen3-TTS clone a voice like ElevenLabs?
Yes. The Qwen3-TTS-12Hz-1.7B-Base model clones a voice from a reference audio clip plus transcript, with a 3-second clip cited in the model card. ElevenLabs offers Instant Voice Cloning from the Starter plan onward.
Related reading
ElevenLabs tool profile, ElevenLabs alternatives radar guide, Descript tool profile, All radar posts