step-3.7-flash

Description

StepFun step-3.7-flash — sparse MoE with 198B total and 11B active parameters, native 256K context, and native understanding of images and video alongside text. Reasoning depth is selectable per request through reasoning_effort (low / medium / high), and the model supports reliable multi-step tool calling, JSON Schema structured output, streaming, and prompt caching. Tuned for agent and coding workloads. On /v1/responses, send the whole conversation in the input array on every turn: no conversation state is kept upstream, so requests that carry previous_response_id or conversation are refused.