Nari Qwen3 Tts

Nari Labs offers optimized multimodal model APIs, starting with speech. Its public Model APIs serve Qwen3 TTS for text to speech and Qwen3 ASR for speech to text, each in a Standard and a Fast variant. Fast variants are tuned for lower latency, with the vendor citing server side time to first audio under 50 ms for text to speech and under 40 ms to first text for transcription. The Qwen3 models are open source models that Nari serves on its own inference stack. Beyond the public APIs, Nari sells dedicated inference on managed, dedicated or private infrastructure, with kernels, batching, quantization and routing tuned to a customer's workload, and private training to fine tune a model for a specific use case. The company is also behind the open source Dia dialogue text to speech model series and an in house avatar system called Narvatar, listed as coming next. The audience is developers shipping realtime voice products who want lower latency and cost than larger speech vendors. Pricing is usage based with no platform subscription. Qwen3 TTS Standard is $5.00 per 1M input characters and Fast is $10.00. Qwen3 ASR Standard is $0.06 per input audio hour and Fast is $0.12. Accounts top up from $10 for 3 member seats and 2 concurrent requests per model, and a single purchase of $100 or more unlocks 10 member seats and 20 concurrent requests for each of text to speech and speech to text. Larger volumes are handled through sales.