StepZenith officially announced on March 4, following the open-sourcing of the Step3.5Flash model, further achieving full-stack open-sourcing of the Agent base model, and offering pre-trained weights (Base), mid-training weights (Midtrain), and the accompanying Steptron training framework to global developers.

As a model designed specifically for agent scenarios, Step3.5Flash uses a sparse MoE architecture, with a total parameter count of 196 billion. By optimizing to activate approximately 11 billion parameters only during inference, it achieves an extremely high energy efficiency ratio. In single request code tasks, its inference speed can reach up to 350 TPS. With outstanding complex reasoning capabilities and long-chain task handling, the model now has the ability to challenge top closed-source models in terms of inference depth.

QQ20260304-114851.png

Currently, Step3.5Flash is active in the open-source community, with its download count on Hugging Face exceeding 300,000 and ranking first on OpenRouter Trending. In the well-known open-source project OpenClaw, known as "the crayfish," the model has climbed to the second position globally in terms of usage volume, thanks to its significant advantages in speed, stability, and Agent compatibility.

This full-stack open-sourcing not only enriches the diversity of the open-source large model ecosystem but also provides developers with more flexible and transparent underlying support for building high-performance Agents, marking an important breakthrough for domestic large models in balancing model performance and inference costs.