Chat with Step 3.7 Flash online

Step 3.7 Flash is StepFun's efficiency model, released on 28 May 2026 under an Apache 2.0 licence. It is a sparse mixture-of-experts design of about 198 billion parameters that activates only around 11 billion per token, which is the whole point: it answers quickly and costs little, while staying capable enough for real work.

StepFun built it for coding agents and search-style workflows — the kind of job where a model is called many times in a loop and both latency and price compound. Throughput runs to several hundred tokens a second on a good rail, and the model exposes selectable reasoning levels so a caller can trade depth against speed per request rather than being locked into one setting. It takes tools and tool_choice for function calling and supports a JSON response format when the output feeds a pipeline.

The value claim is the interesting part. On SWE-Bench Verified with its advisor mode enabled, StepFun reports Step 3.7 Flash reaching around 97% of a frontier model's coding score at roughly a ninth of the per-task cost. Treat any single benchmark carefully, but the shape of the claim matches what the architecture is for.

The context window is 262,144 tokens, so a substantial codebase or a long document set fits in one conversation. On NinjaChat it is included in every plan — a good default when you want a capable model to answer fast and often.