Cerebras and HF bring Gemma 4 to voice AI

Hugging Face and Cerebras published a speech-to-speech pipeline on June 30 that chains three models: Parakeet for speech-to-text input, Gemma 4 running on Cerebras hardware for reasoning, and Qwen3TTS for text-to-speech output. The pipeline runs at real-time latency. Audio goes in, audio comes out without a perceptible gap between them.

The system is already deployed on Reachy Mini, Hugging Face’s open-source robot platform. More than 9,000 Reachy Mini units are in the field, giving the pipeline an immediate hardware base and a real-world test across a range of environments.

The latency result matters because it breaks a tradeoff that has constrained voice AI for years. Most live voice systems route through smaller distilled models specifically to stay under the latency ceiling. Using a full reasoning model like Gemma 4 in the middle of the loop would, until recently, have made the delay obvious to the person talking. Cerebras’s wafer-scale silicon closes that gap: inference is fast enough to run the reasoning step without breaking the conversational flow.

For teams building voice-first products, the relevant signal is that the pipeline is open-source and the components are individually swappable. A team already using Gemma 4 for text tasks can extend it to voice by adding Parakeet and Qwen3TTS at the edges.

The Devin Fusion release covered earlier this week uses similar logic about model routing under latency constraints: run the capable model where it counts, a cheaper one where it does not. For voice AI specifically, the reasoning depth guide covers how to think about what reasoning power a given task actually needs.