Step 3.5 Flash

Step 3.5 Flash is a sparse Mixture-of-Experts foundation model that activates roughly 11B of its 196B total parameters per token, a design StepFun markets as “intelligence density” that lets it approach the reasoning depth of larger proprietary systems while keeping inference costs manageable. The model is delivered as an open-weight release under Apache 2.0, which makes it attractive for teams that want to self-host or fine-tune without depending on a closed API, and it is exposed through both the Step Plan hosted endpoint and third-party channels such as NVIDIA NIM and OpenRouter.</para>

The release is positioned for agentic and developer workflows where latency and reliable tool use matter as much as raw answer quality. NVIDIA NIM listings cite real-time throughput in the 100–300 tok/s range, peaking around 350 tok/s for coding tasks, and the Step Plan documentation explicitly recommends it for agent and coding scenarios alongside higher-effort reasoning variants. A follow-up Step 3.5 Flash 2603 refresh tuned this profile further toward high-frequency agent use with improved token efficiency and a switchable low-inference mode, suggesting the original checkpoint is intended as a balanced general-purpose option within that line.</para>