198B total / ~11B active (196B language + 1.8B vision)
256K tokens
Apache 2.0
Official open-weight sparse MoE vision-language release positioned as a high-efficiency production model for real-world agents. StepFun describes Step 3.7 Flash as a multimodal model with a 196B language backbone plus a 1.8B vision encoder, activating about 11B parameters per token, with native image understanding, 256K context, and three selectable reasoning levels: low, medium, and high. The public model card emphasizes high-frequency agent workloads combining perception, search, reasoning, and tool orchestration, with claims of up to 400 tokens per second throughput and deployment support across vLLM, SGLang, Transformers, llama.cpp, NVIDIA NIM, and related infrastructure. Benchmark tables in the official materials highlight strong multimodal and agentic performance, including first place on SimpleVQA (Search), 67.1 on ClawEval-1.1, 49.5 on Toolathlon, 56.3 on SWE-Bench Pro, and 59.5 on Terminal-Bench 2.1. StepFun also presents the model as compatible with mainstream agent harnesses and suitable for multimodal enterprise workflows such as financial document parsing, visual search, and concurrent coding agents.