Search
Following the open-sourcing of the Step 3.5 Flash model, Jieyue Xingchen, a Chinese startup specializing in large-scale AI models, has taken a further step by open-sourcing the pre-training weights and intermediate training weights of its Agent foundational model. Additionally, the company has also released the accompanying Steptron training framework. It is noteworthy that the Step 3.5 Flash model employs a sparse Mixture of Experts (MoE) architecture, boasting a staggering 196 billion parameters. However, during the inference process, only around 11 billion parameters are activated. When handling single-request code tasks, the model's inference speed can soar up to an impressive 350 Transactions Per Second (TPS), showcasing its remarkable efficiency and performance.
-
C114 Communication Network -
Communication Home
