StepFun opens Step 3.5 Flash base weights, training code and SFT data

StepFun released the full pipeline behind Step 3.5 Flash rather than only the finished chat model: a pretrained base checkpoint, a second midtrain checkpoint built for code, agents and long-context work, and SteptronOSS, the open-sourced training code. A base checkpoint is the model before it has been tuned to follow instructions or chat, so releasing it lets outside teams run their own post-training rather than start from, and be limited by, the choices StepFun made when tuning the final chat model.

In follow-up posts over the following weeks, StepFun also published the supervised fine-tuning (SFT) dataset used to tune the model, first on Hugging Face and later mirrored to ModelScope. Releasing base weights, training code and SFT data together is a broader disclosure than most open-weight releases, which typically stop at the final checkpoint; this is StepFun's own description of what it released, without independent confirmation of dataset completeness.