StepFun released its 600B parameter model as open weights on October 14. Developers can now self-host this tool for coding and document analysis tasks.
Key points
- Step 5 Preview uses 600B total parameters with only 27B active per token.
- The model supports up to one million tokens of context window.
- Checkpoints are available in FP8 and BF16 formats on Hugging Face.
StepFun released the open weights for its Step 5 Preview model on October 14. This allows developers to self-host the system for complex tasks like coding and finance workflows.
What happened
StepFun published the Step 5 Preview model as open weights on October 14, according to MarkTechPost. The company previously launched an API for the model on September 20. The weights are available on Hugging Face and ModelScope in both FP8 and BF16 formats. The model is designed for long-horizon agentic work, such as navigating large codebases and analyzing documents. It contains 600 billion total parameters but activates only 27 billion per token. This sparse Mixture-of-Experts design allows it to handle up to one million tokens of context.
Running the model requires significant hardware. The FP8 version needs four B300 or eight H200 GPUs. The BF16 version requires eight B300 or sixteen H200 GPUs. Performance tests show it scores 88.3% on long-context reasoning benchmarks. It trails behind closed models like GPT-6 Astra and Claude Opus 5. On coding tasks, it performs near peers like GLM-5.3 and Kimi K3. However, it scored poorly on system reliability benchmarks, achieving only 0.4% on the SRE Bench.
Why it matters
The fix
- Verify that your hardware meets the GPU requirements before attempting to download.
- Check the model's performance on specific tasks relevant to your workflow.
- Monitor the model's behavior during long-horizon agentic tasks for errors.
- Review the license terms associated with the open weights release.
Why it matters
Teams evaluating open-weight models should understand the difference between total and active parameters. A high total count does not mean high compute cost if the model is sparse. Users must also verify hardware compatibility before deployment. Context window size determines how much data the model can process at once. Large windows help with long documents but require more memory. Benchmark scores vary by task type. A model strong in coding may be weak in system reliability. Review multiple metrics to assess true capability. Do not assume performance on one benchmark applies to all use cases.
What to do and how to stay safe: StepFun
- Audit your current GPU inventory to see if it supports FP8 or BF16 inference.
- Test the model on a small subset of your specific workloads before full deployment.
- Monitor resource usage closely during initial runs to prevent system overload.
- Keep an eye on official channels for future updates or patches to the model.
Step-by-step guide: What is password policies and how they protect systems
General security guidance from the Crypt Weekly newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How many parameters does Step 5 Preview have?
The model has 600 billion total parameters, with 27 billion active per token.
What hardware is needed to run the FP8 version?
Running the FP8 version requires either four B300 GPUs or eight H200 GPUs.
How does it perform on coding benchmarks?
It performs near open-weight peers like GLM-5.3 and Kimi K3 on coding tasks.
