Step 3.5 Flash (static.stepfun.com)

Step 3.5 Flash has been unveiled as a highly advanced open-source foundation model designed for superior reasoning and agentic capabilities with remarkable efficiency. Utilizing a sparse Mixture of Experts (MoE) architecture, it selectively activates 11 billion of its 196 billion parameters per token, allowing it to match the reasoning depth of some leading proprietary models while ensuring agility for real-time tasks. With a generation throughput of 100–300 tokens per second, powered by a 3-way Multi-Token Prediction mechanism, this model excels in complex reasoning tasks, showcasing reliability in various real-world scenarios. The significance of Step 3.5 Flash lies in its ability to integrate adaptive reasoning and dynamic tool-use, transforming it into an active agent rather than a mere predictive tool. Its architecture supports a cost-efficient 256K context window, enabling seamless handling of extensive datasets and user-defined tasks. Furthermore, the model's performance in critical benchmarks like SWE-bench and its efficiency in executing Python code for agentic tasks denote a step forward in autonomous coding and professional data analysis capabilities. This positions Step 3.5 Flash as an innovative solution for developers and researchers, facilitating end-to-end problem-solving in real-time applications while prioritizing data privacy and resource management.