Step 3.5 Flash

Step 3.5 Flash is StepFun's February 2026 open-weight foundation model. It uses a sparse mixture-of-experts architecture with 196B total parameters but only 11B active per token, a small-sparse design that the company says beats larger dense rivals on reasoning and coding benchmarks while serving at a fraction of the compute cost.

StepFun reports 97.3% on AIME 2025 and 74.4% on SWE-bench Verified, alongside strong agentic scores. The model was trained with tool use and multi-step coherence as first-class objectives, and ships with 256K context. At 11B active parameters it can serve from a single H100-class GPU with high throughput.

It is licensed Apache 2.0, including training code released as SteptronOss. StepFun later skipped a Step 4 entirely and moved to Step 5 Preview. Benchmark claims are vendor-reported and should be read with the usual skepticism.