Hugging Face and Cerebras Release Real-Time Gemma 4 Speech Model

On July 1, 2026, Hugging Face and Cerebras launched a new speech AI: a voice system built on Google DeepMind's Gemma 4 language model, with 31 billion parameters and predictable real-time latency.

What the system is made of

The architecture has four independent components, each of which can be swapped without rebuilding the others:

• Speech recognition: Nvidia Parakeet
• Language model: Gemma 4 from Google DeepMind (31 billion parameters)
• Inference platform: Cerebras
• Speech synthesis: Qwen3TTS from Alibaba

This approach differs fundamentally from monolithic voice pipelines: when a more accurate ASR model or a faster TTS engine comes out, it can be replaced in the pipeline without shutting down the whole system. That matters especially in the fast-moving field of open voice models.

Developers get an interactive demo in the Hugging Face Space "HF Realtime Voice" and the full source code in the huggingface/speech-to-speech repository on GitHub. Any of the four layers can be forked and adapted to a specific task, from a robotic assistant to an enterprise call center.

The Hugging Face and Cerebras partnership is part of a broader trend: inference speed is becoming as much of a competitive advantage as the quality of the base model. For the open-source ecosystem, this means low latency is no longer the exclusive privilege of closed APIs.

Why does 95th-percentile latency matter?

Median latency stopped being a measure of quality long ago: most commercial voice systems manage an acceptable 300–500 ms on average. The real problem is the 95th percentile (P95), where multi-second pauses appear that users perceive as the other party freezing.

The situation gets worse in multi-turn dialogues, when the model has to call an external tool, process an image or stitch together several context fragments. Each extra step multiplies the delay, and P95 becomes the architecture's Achilles' heel. Cerebras speeds up Gemma 4 inference enough that tail latencies become predictable, so a system can be built with firm response guarantees.

Real-world scale backs up the argument: more than 9,000 Reachy Mini robots already run in production on Hugging Face's speech-to-speech pipeline. Deployments like these expose the gap between lab benchmarks and real operating latency.

What this means

An open stack built on Gemma 4 with Cerebras inference lowers the barrier for teams that need voice AI without proprietary dependencies. Modularity preserves long-term flexibility: each of the four layers can be updated independently as better models ship, so there is no need to rewrite the whole pipeline for a single improvement. The public demo and open repository turn the concept into a production-ready template for developers of robotics, smart devices and voice interfaces.

Frequently asked questions

How many parameters does Gemma 4 have in this pipeline?

The pipeline uses the 31-billion-parameter version of Gemma 4 from Google DeepMind. Inference runs on the Cerebras platform, which delivers predictable latency even at the 95th percentile of load.

Where can I try the system?

The demo is available in the Hugging Face Space "HF Realtime Voice", and the full source code is open in the huggingface/speech-to-speech repository on GitHub.

What is Gemma 4?

It is a language model from Google DeepMind with 31 billion parameters, used in the open speech-to-speech pipeline for real-time voice AI.

Can Gemma 4 be used for free?

Yes, this is the first publicly available modular stack for voice AI, and developers can use and replace each of its components independently.

What is the Gemma 4 speech model?

It is an open voice AI launched by Hugging Face and Cerebras on top of Google DeepMind's Gemma 4 language model (31 billion parameters). The system converts speech to speech with predictable real-time latency, using a modular stack of four independent components.