Search
On March 4, StepFun took a significant leap by fully open-sourcing the Step 3.5 Flash model. This release included not only the model itself but also its pre-training weights, intermediate training weights, and a comprehensive training framework to support its deployment. The model stands out for its innovative sparse Mixture of Experts (MoE) architecture, boasting an impressive total of 196 billion parameters. However, during the inference process, it cleverly activates only about 11 billion parameters, optimizing efficiency. For single-request code tasks, it achieves a remarkable maximum inference speed of 350 TPS (Transactions Per Second). Presently, the model has garnered significant attention, with over 300,000 downloads on Hugging Face, and has surged to the forefront, claiming the top spot in terms of usage on OpenClaw on a global scale.
-
C114 Communication Network -
Communication Home
