Cerebras adds support for Google's Gemma 4, running the 31B multimodal model at 1,500 tokens per second · Digg