StepFun Releases Step 5 Preview Open Weights for Agentic Work

StepFun released the weights for its Step 5 Preview model on October 14. The company made FP8 and BF16 checkpoints available on Hugging Face and ModelScope, following an API launch on September 20, according to MarkTechPost.

The Step 5 Preview is a sparse Mixture-of-Experts model designed for long-horizon agentic work. It contains 600 billion total parameters but activates only 27 billion per token. The architecture supports up to one million tokens of context, offering developers a self-hostable option for complex tasks.

Technical Specifications and Performance

StepFun’s engineering team describes the model as having 92 transformer blocks in a narrow, deep design. It uses a mix of gated sliding-window attention and full attention layers, interleaved in a three-to-one ratio. Sparse grouped-query attention allows full-attention layers to pick 4,096 history tokens per query.

An MTP-3 head enables speculative decoding, providing roughly a 30 percent end-to-end speedup. Training utilized FP8 mixed precision and reinforcement learning on million-level verifiable agentic environments. The model runs only on data-center GPUs, requiring four B300s or eight H200s for FP8.

Benchmarks show mixed results compared to peers. Step 5 Preview scored 88.3 percent on the AA-LCR long-context reasoning test, according to Artificial Analysis. It trailed GPT-6 Astra and Claude Opus 5 but performed near open-weight rivals GLM-5.3 and Kimi K3 on coding tasks.

Deployment Requirements and Use Cases

The model is positioned for software engineering, professional knowledge work, and finance. It can navigate large codebases, coordinate multi-file changes, and run code iterations. It also analyzes long documents, charts, screenshots, and video to produce structured reports.

BF16 deployment requires eight B300s or 16 H200s. Performance varies by domain; it achieved 66.4 percent on FrontierFinance, beating GLM-5.3 but trailing Claude Opus 5. Weak spots include 0.4 percent on SRE Bench and 9.2 percent on Harvey’s Legal Agent Benchmark, per Vals AI.

What should you do?

Organizations evaluating large language models for agentic workflows should verify hardware compatibility before deploying Step 5 Preview. The model demands significant GPU resources, so infrastructure planning is essential. Teams should also review benchmark scores relevant to their specific industry verticals.

Background: Large language models

Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.

Read the full guide: Large Language Models vs Generative AI: The Distinction and Security Implications

What to do and how to stay safe: StepFun

• Audit existing GPU inventory to determine if it meets the B300 or H200 requirements for FP8 or BF16 inference.
• Evaluate the model on internal datasets to verify performance on specific coding or document analysis tasks.
• Monitor community feedback on Hugging Face and ModelScope for stability issues in early open-weight releases.
• Compare cost-per-token against API alternatives if self-hosting infrastructure is not currently available.

Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

When were the Step 5 Preview weights released?

The weights were released on October 14, following an API launch on September 20.

What hardware is required to run Step 5 Preview?

FP8 requires four B300s or eight H200s; BF16 requires eight B300s or 16 H200s.

How does Step 5 Preview compare to other models?

It performs near GLM-5.3 and Kimi K3 on coding but trails GPT-6 Astra and Claude Opus 5.