Open-weight 198B sparse MoE with 11B active parameters, native image input, selectable reasoning effort, and agent-focused coding performance.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Agentic Coding Gains Over Step 3.5 Flash at Flash-Class Speed
Benchmarks
Artificial Analysis: Speed-Intelligence Position and a Factuality Gap
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Comparison
Step 3.7 Flash or Step 3.5 Flash: What Does the Upgrade Buy?
Step 3.5 Flash is the direct predecessor, built on the same 196B-parameter sparse MoE language backbone with 11B active parameters and a 256K context window. Step 3.7 Flash adds a vision encoder and higher agentic scores: in StepFun's results, SWE-Bench Pro rises from 51.3 to 56.3 and Terminal-Bench 2.1 from 53.4 to 59.5. The practical decision is whether those gains and image input justify a substantially higher output price.
Choose Step 3.7 Flash when you need image or screenshot input, stronger agent results (ClawEval-1.1: 67.1 vs 43.6), or steadier results across coding harnesses. Choose Step 3.5 Flash when workloads are text-only and high-volume, since its output price is markedly lower, or when you run Claude Code-style harnesses, where StepFun's own Step-SWE-Bench scores show 3.5 Flash slightly ahead (73.0% vs 71.5%).
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Coding agents
StepFun reports 56.3 on SWE-Bench Pro and 59.5 on Terminal-Bench 2.1, and its Step-SWE-Bench results across six harnesses, including Claude Code, KiloCode and OpenCode, range from 64.5% to 71.5%. That consistency suits agents that edit files and run terminals in loops. These are provider-reported numbers; for the hardest repository tasks, Step 5 Preview scores higher on Terminal-Bench 2.1.
Screenshot-to-code and GUI tasks
A 1.8B-parameter vision encoder lets the model read UI screenshots, charts and documents directly, and NVIDIA lists GUI-oriented tasks and frontend generation among its intended uses. Artificial Analysis measured 75.3% on MMMU-Pro. It fits frontend generation from mockups and UI-testing agents. Output is text only, so it can describe or code what it sees but cannot generate images.
Search and research agents
StepFun reports 67.1 on ClawEval-1.1 and 49.5 on Toolathlon, and BenchLM lists 75.8% on BrowseComp, pointing to dependable multi-step tool use. That suits agents that search, cross-check sources and compile answers. Pair it with retrieval, because its measured hallucination rate means ungrounded recall is a weak point.
High-frequency calls with tunable effort
Reasoning effort can be set per request: low for extraction, summarization and rewriting, high for planning and math. With about 11B active parameters and roughly 400 tokens/s reported at launch on StepFun's API, it suits pipelines that mix simple and hard steps. Measure real speed on your access path, since later figures are lower.