Google's Gemma 4 31B Multimodal Model Now Runs on Cerebras at Over 1,800 Tokens Per Second

Google DeepMind's open-weight Gemma 4 31B model is now available in public preview on Cerebras Inference, where it runs at more than 1,800 tokens per second—a speed the company bills as the fastest multimodal inference available today. The launch pairs one of Google's most capable open models with Cerebras's wafer-scale hardware, promising near-instant responses for real-time visual and agentic workloads.