Skip to main content
Hopper serves three models, one for each stage of the voice agent pipeline. Request a model by its API ID on the endpoints listed in its section below.

​
Gemma 4 31B

Hopper serves Gemma 4 31B, Google’s open model, as NVIDIA’s NVFP4-quantized checkpoint nvidia/Gemma-4-31B-IT-NVFP4. It handles chat completions with streaming and tool calling on the OpenAI-compatible API. Time to first token is ~37 ms on short prompts, measured 2026-08-02 from us-west-2 with cold, unique prompts. See Chat completions for request parameters, streaming, and tool calling.

​
OmniVoice

Hopper serves OmniVoice, k2-fsa’s open text to speech model (k2-fsa/OmniVoice). It generates 24 kHz English audio and streams as WAV or raw PCM; MP3 is available for non-streaming requests. Voices come from the shared library or from cloning: instant cloning uses a single clip of at least 10 seconds, and pro cloning trains on a larger dataset asynchronously. See Generate speech for request parameters, and Voice cloning for instant and pro cloning.

​
Nemotron ASR

Hopper serves Nemotron ASR, NVIDIA’s open streaming speech-to-text model (nvidia/nemotron-3.5-asr-streaming-0.6b). It transcribes uploaded audio over HTTP and live audio over WebSocket, returns word-level timestamps, and accepts keyterm lists to bias recognition. Streaming sessions include turn detection with three presets: balanced, patient, and responsive. See Transcription for batch requests and Streaming for the WebSocket protocol.