Hugging Face and Cerebras bring Gemma 4 to real-time voice AI

Hugging Face and Cerebras have partnered to create a real-time voice AI with significantly reduced latency, making conversations feel more natural. The solution is an open, cascaded speech-to-speech pipeline that uses Nvidia's Parakeet for speech recognition, Google's Gemma 4 for language model inference, and Alibaba's Qwen3TTS for text-to-speech. Cerebras hardware dramatically accelerates the language model inference step, providing stable, low-latency responses that are crucial for real-world interactions. This technology is already being applied in robotics, such as the Reachy Mini robots, where responsiveness is essential. The collaboration demonstrates how open-source models paired with high-performance hardware can advance the next generation of conversational AI.