Step 3.5 Flash Open Source: StepFun Uses Sparse Models to Compete for Efficient Agent Reasoning
StepFun open-sourced Step 3.5 Flash in early 2026, positioning it as a foundation model balancing reasoning, tool use, and efficiency.
StepFun moves from a multimodal product company into the global open model competition, attempting to differentiate with small activated parameters, long context, and agent capabilities.
The official repository reveals the model is designed for long-chain reasoning and agent tasks, offering open weights, inference instructions, and integration with mainstream frameworks.
The domestic model market is no longer dominated solely by top-tier giants and a few star companies; more teams are gaining developer attention and overseas visibility through open weights.
Technical teams can use it as a candidate for multi-model routing, but need to compare unit effective result cost under the same hardware and tasks.
Observe real tool tasks, inference throughput, overseas community adoption, and continuous updates to StepFun's API products.