StepFun Open-Sources 600-Billion-Parameter Step 5 Model for 24-Hour Agent Tasks

StepFun has released the open-weights version of Step 5 Preview, its 600-billion-parameter sparse Mixture-of-Experts model designed to handle extended autonomous agent execution spanning up to 24 hours, with NVIDIA confirming the NVFP4 quantized variant achieves nearly 8,000 tokens per second per GPU on its GB300 platform. The Shanghai-based AI laboratory introduced the model's API in September 2026, followed by the open-weights release on October 15 under the StepFun Model License 1.0, featuring three checkpoints in BF16, FP8, and NVFP4 formats for flexible deployment.