Model-specific runtimes
Compilation, scheduling, batching, and memory use optimized for each model.
Model APIs
Optimized model APIs
Optimized multimodal model endpoints built for low latency and high efficiency: starting with speech.
Models
Nari Qwen3-TTS 1.7B Fast
Sub 50 ms server-side TTFA$10 / 1M characters
Text to SpeechPublicNari Qwen3-TTS 1.7B
$5 / 1M characters
Text to SpeechPublicNari Qwen3-ASR 1.7B Fast
Sub 40 ms server-side TTFS$0.12 / hour
Speech to TextPublicNari Qwen3-ASR 1.7B
$0.06 / hour
Speech to TextPublicPyannote Community-1
3600x Aggregate RTFx
Speaker DiarizationComing SoonNeed another model?
Share the next model we should optimize and add to our endpoint collection.
Specialized runtimes for multimodal models
Multimodal models need different execution, streaming, and scheduling strategies. We tune the runtime around each model's behavior.
Streaming-native execution
Input and output streaming designed for interactive multimodal workloads.
cURL
Qwen3-TTS
curl --fail --silent --show-error \
https://api.narilabs.com/v1/audio/speech \
-H "Authorization: Bearer $NARI_API_KEY" \
-H "Content-Type: application/json" \
--data '{
"model": "qwen3-tts-fast",
"input": "Multimodal models, ready to ship.",
"response_format": "wav"
}' \
--output speech.wavEvaluate performance for your workload
Review latency, quality, and serving cost together.
Qwen3-TTS
The world’s fastest TTS endpoint, with expressive voices scoring above the field average in preliminary independent benchmarks.
50 ms
Fast: time to first audio
$5
Standard: per 1M characters
$10
Fast: per 1M characters
Expressive voices
Scoring above the field average in preliminary independent benchmarks.
Based on preliminary independent measurements of publicly available TTS endpoints from US East. Results may vary by region, connection, and workload. Pricing checked Sep 8, 2026 from Artificial Analysis.
Start with an optimized model
Choose a model and test it against your workload.