[Step 3.5 Flash Open-Source Model Released by Starlight Step] Jin10 News, February 2 – Starlight Step has released a new generation of open-source Agent base model, Step 3.5 Flash. This model is designed for real-time Agent workflow scenarios, with a maximum inference speed of up to 350 tokens per second. Reportedly, Step 3.5 Flash adopts a sparse MoE architecture, activating approximately 11 billion parameters per token (out of a total of 96 billion parameters). Multiple chip manufacturers, including Huawei Ascend, Muxi Integrated Circuit, Biren Technology, Iluvatar CoreX, Tianshu Zhixin, and Alibaba T-Head, have completed adaptation for the model.