Step 3.5 Flash (Deep Infra)

Step 3.5 Flash is an open-weights reasoning model released by StepFun in February 2026, designed to balance fast response times with the ability to handle structured reasoning and tool-augmented workflows. Its architecture uses a sparse Mixture of Experts design with 196 billion total parameters and only 11 billion active per inference, a configuration that aims to keep compute and latency low while preserving the capacity of a much larger model. The model accepts text input and produces text output, and it is positioned for developers who need reasoning behavior, tool calling, temperature control, and structured output in a single text-generation endpoint.

The repository published under the StepFun GitHub organization documents additional architectural details, including multi-token prediction handling and the context and attention scheme used by Step 3.5 Flash, giving integrators a clear view of how the model is built beyond the headline parameter counts. Open weights are available alongside cloud-hosted access, which makes the model practical both for teams that want to self-host and for those that prefer a managed inference path. Together, the sparse MoE footprint and the documented training-time techniques make Step 3.5 Flash a reasonable fit for latency-sensitive assistant and agent workloads that still benefit from explicit reasoning and structured tool use.