阶跃星辰发布开源基座模型 Step 3.5 Flash 多家头部芯片厂商完成适配

Core Insights - The core focus of the news is the launch of the new open-source Agent base model, Step 3.5 Flash, by Jiyue Xingchen, which is designed for real-time Agent workflow scenarios, balancing inference speed, intelligence level, and cost efficiency [1][3] Model Performance - Step 3.5 Flash achieves a maximum inference speed of 350 tokens per second for single request code tasks [1][3] - The model utilizes a sparse MoE architecture, activating approximately 11 billion parameters per token out of a total of 196 billion parameters, significantly enhancing inference efficiency while maintaining model capability [1][3] Industry Collaboration - Jiyue Xingchen has collaborated with several chip manufacturers, including Huawei Ascend, Muxi Co., Birun Technology, Suiyuan Technology, Tianshu Zhixin, and Alibaba Pingtouge, to adapt Step 3.5 Flash, enhancing model adaptability and computational efficiency [1][3] - The collaboration aims to lower inference costs and reduce the overall barriers for enterprises and developers in applying large models, accelerating the practical application of large models in various scenarios [1][3] Ecosystem Development - In July 2025, Jiyue Xingchen initiated the "MoCore Ecological Innovation Alliance" with nearly 10 chip and infrastructure manufacturers to bridge the technical barriers between chips, models, and platforms [2][4] - The alliance aims to optimize and enhance computational efficiency, facilitating the large-scale application of models across various industry scenarios [2][4] - The industry recognizes that deep collaboration between models and computational resources will be a crucial path for promoting the large-scale application of inference models [2][4]