Gepard : 0.6B streaming TTS built for real-time dialogue - 20× realtime factor, ~50ms time-to-first-audio, vLLM-native, Apache 2.0

We just open-sourced Gepard 1.0, a TTS model built for real-time conversation. It’s streaming-first: audio starts the moment text arrives, generated frame by frame instead of waiting for a full sentence.

**- ~555M params**: Qwen3.5 0.8B backbone (14 layers) + Nemo NanoCodec (FSQ, 22.05kHz)
 **- ~20 x RTF**, **~50ms TTFA** on one RTX 5090 via vLLM
 **- Up to 256 parallel sequences** on a single RTX Pro 6000 Balckwell with 96GB VRAM
 **- Zero-shot voice cloning** from a few seconds of reference
 - Languages: **English (US/UK), Spanish (MX), Portuguese (BR), Dutch**
 **- Apache 2**

Benchmarks (Seed-TTS-eval): we put it head-to-head against VoxCPM2, Fish-S2, OmniVoice, Qwen3-TTS, Echo-TTS, and Chatterbox Turbo on identical texts. Gepard leads the field on perceived quality - top NISQA-MOS (4.25), and cleanest on noise, coloration, and discontinuity.

Honest tradeoff: The streaming-first design costs us on speaker similarity (SIM 0.585) and WER (0.036), so it’s a strong fit where a natural realtime voice matters more than exact voice-matching.

Also you can check how it works on vLLM on our website: https://www.nineninesix.ai