Core Insights - The company, Jieyue Xingchen, has initiated the training of its next-generation model, Step 4, and is inviting developers to participate in its co-development [1][3] - Jieyue Xingchen has released and open-sourced its latest base model, Step 3.5 Flash, which utilizes a sparse mixture of experts (MoE) architecture with a total parameter count of 196 billion, activating approximately 11 billion parameters per token [1][3] - Step 3.5 Flash has been optimized for high inference speed, achieving up to 350 tokens per second for single-request code tasks, addressing the need for low-latency responses in multi-step reasoning scenarios [1][3] Model Performance - Step 3.5 Flash has demonstrated strong performance in various tests, including τ²-Bench, BrowseComp, LiveCodeBench V6, and AIME 2025 mathematics competition, showcasing its competitiveness in agent scenarios and mathematical tasks [2][4] - The model exhibits robust capabilities in automatic programming and "edge-cloud collaboration," enabling the construction of complex visual geographic systems based on natural language instructions and efficiently breaking down complex user requirements into sub-tasks for cross-platform data analysis and decision support [2][4] Industry Trends - There is a noticeable shift among developers from simple prompt engineering to building agents and workflows, indicating a growing interest in more complex AI applications [1][3] - Despite the strength of underlying models, challenges remain in stability, response speed, and cost in scaled scenarios, which Step 3.5 Flash aims to address through its specialized optimizations [1][3]