Introducing Step 3.7 Flash: multimodal reasoning at scale

What this post added

This post introduces the integration of Step 3.7 Flash, a 198-billion-parameter sparse MoE vision-language model, into the Baseten Model Library. It details the hardware-efficient deployment strategy using FP8 quantization, enabling the model to be served on 4xH100s instead of the previously recommended 8xH200s/B200s. This optimization significantly reduces cost per token and improves autoscaling. The post also highlights the architectural features of Step 3.7 Flash, including its sparse MoE design, multimodal input capabilities, long context window, and flexible reasoning, and explains how these, combined with hybrid attention and speculative decoding, allow for efficient production traffic serving on the optimized hardware.