Step 3.7 Flash

Step 3.7 Flash – Sparse MoE Vision-Language Model for Agentic Workflows

Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts vision-language model from StepFun, built on the Step 3.5 Flash language backbone with a dedicated vision encoder added for native multimodal understanding. Its defining design target is high-frequency production agentic workloads that combine perception, search, and reasoning in a single model without requiring a separate vision module inside agent frameworks.

Multimodal Architecture

Step 3.5 Flash was text-only. Step 3.7 Flash introduces native image understanding by pairing the same language backbone with a 1.8B ViT encoder a structural addition rather than a fine-tune. StepFun documents an observed emergent behavior during testing: the model combined visual tools with non-visual tools without being explicitly trained to do so, such as rendering generated frontend code in a GUI and inspecting the result before iterating.

Visual Search Design

For recognition tasks where parametric knowledge is insufficient such as long-tail entities or recently emerged concepts the model invokes a visual search tool to retrieve and verify. Search is integrated into the reasoning loop rather than treated as a separate add-on.

Deployment

Supports vLLM, SGLang, Hugging Face Transformers, and llama.cpp for inference. Available as an NVIDIA NIM microservice. Local deployment requires at least 128GB unified memory (NVIDIA DGX Station, AMD Ryzen AI Max+ 395, Mac Studio / MacBook Pro).