Step 5 Preview, a 1M-context MoE from StepFun, shows up on OpenRouter
Step 5 Preview: StepFun's 1M-Context MoE Now on OpenRouter
StepFun has introduced its premier model for agentic operations, Step 5 Preview, now accessible via OpenRouter. This model is specifically engineered to handle complex, multi-step professional workflows and massive datasets.
🛠️ Technical Architecture & Design
Step 5 Preview utilizes a sparse Mixture-of-Experts (MoE) framework. This allows the model to maintain a massive knowledge base while remaining computationally efficient during inference.
Parameter Breakdown
The model's scale can be represented by the following ratio:
Core Capabilities
- Software Engineering: High proficiency in navigating and editing large codebases.
- Financial Analysis: Specialized strength in professional knowledge work within the finance sector.
- Agentic Workflows: Designed to use tools and iteratively refine results.
- Massive Context: Supports a window of
1.0Mtokens.
💰 Pricing and Performance Metrics
OpenRouter provides transparent tracking of what users actually pay versus the listed rates. Due to caching and various discounts, the real cost is often higher lower than the sticker price.
Cost Structure
| Metric | Rate (per 1M tokens) |
|---|---|
| Input | $1.00 |
| Output | $2.70 |
| Cache Read | $0.05 |
Operational Benchmarks
The following data reflects the P50 performance across providers:
- Throughput:
63 tps(Tokens per second) - Latency (TTFT):
3.49 s - Uptime (3d):
- Availability (3d):
Note: OpenRouter ensures high reliability by automatically routing requests to the next-best healthy provider if the primary one fails.
🔌 Integration & Implementation
The API is fully OpenAI-compatible, meaning developers can integrate it by simply updating the base URL and the model slug: stepfun/step-5-preview.
Example Request Format
To utilize structured outputs, the model supports a response_format using a JSON schema:
{
"model": "stepfun/step-5-preview",
"messages": [{"role": "user", "content": "Analyze this financial report."}],
"response_format": { "type": "json_object" }
}
🤖 The Agentic Ecosystem
Step 5 Preview is designed to power advanced AI agents. Several open-source projects can leverage this model:
OpenClaw: An agent for messaging apps that manages files, browses the web, and sends emails.
Hermes Agent: A self-improving agent by Nous Research with persistent memory and 40+ built-in tools.
Cline: An IDE-integrated coding agent that autonomously edits files and runs terminal commands.
Kilo Code: A cross-platform coding agent (VS Code, JetBrains, CLI) for accelerated shipping.
🔍 FAQ & Model Comparisons
What exactly is Step 5 Preview? It is StepFun's flagship model for agentic tasks. It excels in software engineering and finance, utilizing a 600B parameter MoE architecture to process vast documents and codebases over multiple iterative steps.
Other StepFun Offerings
StepFun provides other specialized models for different efficiency needs:
1. Step 3.7 Flash
A high-efficiency multimodal MoE model.
- Architecture: 196B total / 11B active parameters.
- Context: 256K tokens.
- Unique Feature: Selectable reasoning levels (High, Medium, Low) to balance cost vs. depth.
- Pricing: 1.15 (Output).
2. Step 3.5 Flash
The most capable open-source foundation model from the suite.
- Architecture: Sparse MoE (11B active / 196B total).
- Focus: A reasoning-centric model optimized for extreme speed, even when handling long contexts.