Step 3.7 Flash is a multimodal sparse Mixture-of-Experts (MoE) model developed by StepFun, designed for high-efficiency production-level agents. Released in May 2026, the model succeeds Step 3.5 Flash by integrating native vision capabilities and optimized reasoning tiers. It is engineered to balance high throughput with complex task execution, specifically targeting agentic coding, multi-step tool orchestration, and long-horizon workflows.
The model's architecture consists of a 196B-parameter language backbone paired with a 1.8B-parameter vision encoder (ViT), totaling approximately 198 billion parameters. However, its MoE design ensures that only roughly 11 billion parameters are activated per token, allowing for inference speeds of up to 400 tokens per second. It further utilizes a 3-way Multi-Token Prediction (MTP) mechanism and hybrid Sliding Window/Global attention to maintain a 256,000-token context window while minimizing memory overhead.
A distinctive feature of Step 3.7 Flash is its selectable reasoning effort system. Developers can choose between three computation tiers—low, medium, and high—to dynamically trade off latency and cost against reasoning depth on a per-call basis. This capability is paired with strong performance in agentic benchmarks, such as a 74.4% score on SWE-bench Verified and 67.1% on ClawEval-1.1, reflecting its proficiency in navigating real-world terminal environments and complex software repositories.
The model provides native support for image and video understanding, enabling it to parse screenshots, documents, and user interfaces without requiring a separate vision module. It is highly compatible with the Model Context Protocol (MCP) and popular agent frameworks, making it suitable for building autonomous coding assistants, enterprise search tools, and multi-tool coordination systems.