Step 5 Preview has 600 billion total parameters and 27 billion active per token. NVIDIA says its NVFP4 version runs at nearly 8,000 tokens per second per GPU on GB300.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [NVIDIA AI](https://x.com/NVIDIAAI/status/2110730497907216588)
Why it matters #
Downloadable weights and NVIDIA's serving claims give developers a way to test Step 5 outside StepFun's API. The model card describes long-horizon tool use, while NVIDIA's 24-hour framing remains a company claim, not independently demonstrated performance.
StepFun's 600-billion-parameter Step 5 Preview weights are available on Hugging Face. NVIDIA says the NVFP4 version runs at nearly 8,000 tokens per second per GPU on its GB300 platform. StepFun introduced the model on September 20th and said weights would follow on October 15th; NVIDIA described the release as StepFun's largest open-weights release in its October 15th post.…