Kokoro 82M vs Qwen3 TTS
Two open-weight text-to-speech models on the same endpoint and API key, so switching between them is a one-word change. Below: what each costs for three realistic workloads, how they differ, and measured speed.
Inference APIs prices are read live from the price list.
- Cost: Kokoro 82M is cheaper in all three workloads, by 80%.
Side by side
| Kokoro 82M | Qwen3 TTS | |
|---|---|---|
| Served by | Inference APIs | Inference APIs |
| Model author | Hexgrad | Alibaba Qwen |
| Model id | hexgrad/Kokoro-82M | Qwen/Qwen3-TTS |
| Price | $5.20 / 1M characters | $26.00 / 1M characters |
| Capabilities | Text to speech, 54 voices, MP3 / WAV, 8 languages | Text to speech, 9 voices, Instruction control, MP3 / WAV, 10 languages |
| Weights | Open | Open |
| Status | Available | Available |
| Limits | No per-minute or per-day request or token caps; usage draws on a prepaid balance | No per-minute or per-day request or token caps; usage draws on a prepaid balance |
| 300 characters → audio | 0.84 s | — |
What three workloads cost
List prices applied to the same work. The cheaper side of each row is in bold.
| Workload | Kokoro 82M | Qwen3 TTS | Difference |
|---|---|---|---|
| One 1,000-word article (about 6,000 characters) | $0.03 | $0.16 | 80% |
| One hour of narration (about 54,000 characters) | $0.28 | $1.40 | 80% |
| One million characters (about 18 hours of speech) | $5.20 | $26.00 | 80% |
Kokoro 82M is cheaper in all three workloads, by 80%.
When to choose which
- Cost is the deciding factor; it is cheaper in every workload above
- Interactive speech: sub-second responses for a paragraph in our measurement
- You need voices in several languages at a low price per character
- Price per character matters most: Kokoro is five times cheaper
- You need English voice variety: Kokoro has 54 voices, this model has 9
- You want to steer tone and emotion with plain-language instructions ("speak slowly and calmly")
- Ten languages from one model, including Korean, Russian and German, which Kokoro does not cover
- Voice cloning from a short sample, which the host documents through a voice_id parameter
- You need emotional delivery or non-verbal sounds such as laughter
Measured speed
Median of three runs on 2026-09-16: a 300-character paragraph to mp3, timed from request to last byte. Expect ±30% with time of day. Method and raw numbers: speed measurements.
Trying both
Same endpoint, same key. Change one string:
for MODEL in "hexgrad/Kokoro-82M" "Qwen/Qwen3-TTS"; do
curl -s https://api.inferenceapis.com/v1/audio/speech \
-H "Authorization: Bearer $INFERENCE_API_KEY" -H "Content-Type: application/json" \
-d "{\"model\": \"$MODEL\", \"input\": \"The quick brown fox jumps over the lazy dog.\"}" \
--output "$(basename $MODEL).mp3"
doneFAQ
Which is cheaper, Kokoro 82M or Qwen3 TTS?
Kokoro 82M is cheaper in all three workloads, by 80%. The table above uses list prices: Kokoro 82M $5.20 / 1M characters; Qwen3 TTS $26.00 / 1M characters.
How do I switch between them?
They are on the same endpoint and key. Change the model field from hexgrad/Kokoro-82M to Qwen/Qwen3-TTS and nothing else.
Spotted a price or limit that has changed? Tell us and we will re-check the provider page.
