Gepard 1.0

Gepard 1.0 is an autoregressive, prosody-aware text-to-speech model designed specifically for real-time conversational AI. It utilizes a Qwen3.5-based backbone to generate audio in single-pass frames, enabling ultra-low latency and high-throughput streaming suitable for voice agents.

Gepard 1.0 is an autoregressive, prosody-aware text-to-speech model designed specifically for real-time conversational AI. It utilizes a Qwen3.5-based backbone to generate audio in single-pass frames, enabling ultra-low latency and high-throughput streaming suitable for voice agents.

• 01Single-pass, streaming-first frame generation for low-latency dialogue
• 02vLLM-native support allowing for high-throughput concurrent sessions
• 03Zero-shot voice cloning integrated into the prefill stage
• 04Classifier-free guidance (CFG) distilled into weights for quality at no extra compute cost
• 05Optimized for real-time voice agents with ~50ms time-to-first-audio