Step 3.5 Flash

Step 3.5 Flash is StepFun's open-source reasoning model built on a sparse Mixture of Experts architecture with 196.81B total parameters and roughly 11B active per token. It delivers high throughput through a 45-layer transformer, 3-way multi-token prediction, and hybrid sliding-window attention, generating 100-300 tokens per second in typical usage. The model excels at agentic coding and software engineering, achieving 74.4% on SWE-bench Verified and 97.3 on AIME 2025. It provides reliable function calling and tool orchestration for multi-step task decomposition. Key strengths include a 256K token context window for long-document processing, Apache 2.0 licensing, and open weights for self-hosting. The model targets a strong speed-to-capability ratio, making it well-suited for real-time agent workflows, logical reasoning, mathematics, deep research, and complex task planning.