StepFun Releases Step 5 Preview as Open Weights: 600B MoE, 27B Active Parameters, 1M-Token Context
StepFun has released Step 5 Preview as open weights, with FP8 and BF16 checkpoints now available on Hugging Face and ModelScope. Step 5 Preview is a sparse Mixture-of-Experts (MoE) model built for long-horizon agentic workloads. It features 600B total parameters, 27B active parameters per token, and up to 1M tokens of context.
For developers, this adds a new self-hostable option for coding agents, document analysis, and finance workflows—areas where long-context reasoning and sustained multi-step execution matter most.
TL;DR
- Size: 600B total parameters, 27B active per token, up to 1M-token context.
- Hardware: Data-center GPUs only. FP8 requires 4x B300 or 8x H200; BF16 requires 8x B300 or 16x H200.
- Performance: Roughly on par with open-weight peers GLM-5.3 and Kimi K3 on coding tasks, but behind GPT-6 Astra and Claude Opus 5.
- Best result: 88.3% on AA-LCR long-context reasoning (Artificial Analysis).
- Weakest result: 33.3% on Terminal-Bench 4.0, compared with 41.9% for GLM-5.3.
What Is Step 5 Preview?
Step 5 Preview is a sparse MoE model from StepFun, designed specifically for long-horizon agentic work. Its 600B total parameter count and 27B active parameters per token place it in the same class as other large open-weight MoE models released in 2026, where the goal is to balance raw capability against inference cost.
The 1M-token context window is the headline feature for agentic use cases. Long-horizon agents—those that plan, execute, and revise across many steps—accumulate large working contexts from tool outputs, codebases, and documents. A 1M-token window reduces the need for aggressive context compression or retrieval scaffolding, though effective long-context reasoning remains a separate challenge from raw context length.
Deployment Requirements
Step 5 Preview is not a model you will run on a workstation. The published requirements are explicit:
| Precision | Minimum Hardware |
|-----------|------------------|
| FP8 | 4x B300 or 8x H200 |
| BF16 | 8x B300 or 16x H200 |
This places Step 5 Preview firmly in data-center territory. Teams evaluating it should budget for multi-GPU nodes and consider FP8 as the default deployment path unless BF16 precision is required for a specific workload.
Benchmark Performance
StepFun's positioning is clear: Step 5 Preview competes with other open-weight frontier models, not with the proprietary leaders.
- Long-context reasoning: 88.3% on AA-LCR (Artificial Analysis), the model's strongest showing.
- Coding and agentic tasks: Competitive with GLM-5.3 and Kimi K3, but behind GPT-6 Astra and Claude Opus 5.
- Terminal-Bench 4.0: 33.3%, versus 41.9% for GLM-5.3—a notable gap for terminal-based agentic workflows.
The Terminal-Bench result is the most important caveat for practitioners. Step 5 Preview is strong at long-context reasoning and broad coding tasks, but teams building command-line or shell-driven agents should validate performance against their own workloads before committing.
Why It Matters
The open-weight model landscape in 2026 has consolidated around sparse MoE architectures, and Step 5 Preview follows that pattern. The trade-off is consistent: very large total parameter counts paired with a much smaller active parameter budget per token, which keeps inference costs manageable while preserving model capacity.
For organizations that cannot or will not route sensitive workloads through proprietary APIs, Step 5 Preview offers a credible self-hostable option for document-heavy analysis, financial workflows, and coding agents. The main constraints are hardware scale and the benchmark gap on terminal-centric agentic tasks.
Getting Started
- FP8 checkpoint: Hugging Face – Step-5-Preview
- BF16 checkpoint: Hugging Face – Step-5-Preview-BF16
- ModelScope mirror: stepfun-ai/Step-5-Preview
Teams should start with FP8 on 8x H200 or 4x B300 nodes to minimize footprint, and benchmark against their own long-context and agentic tasks before scaling deployment.
via MarkTechPost