StepFun’s 600B parameter MoE model, Step5Preview, handles a massive context of 1 million tokens and demonstrates specialized capabilities for agentic tasks.
Imagine this: You have 50 thick accounting reports, each over 1,000 pages, piled up on your desk. What if you asked an AI assistant, “Find and summarize all the anomalies in our company’s financial flow over the past five years”? Previous AI would have required you to input these documents one by one, and even then, it would often lose its place midway. However, ‘Step5Preview’, which recently appeared on OpenRouter, is a new challenger breaking through these impossibilities.
Why is this technology important?
While using AI in daily life, frustrating moments often arise. Sometimes it quickly forgets what you just said, or fails to properly analyze a long document. Experts call this the limitation of AI’s ‘memory’, or its ‘context window’ (the amount of data an AI can process at once).
Step5Preview is equipped with an overwhelming memory capacity of 1 million tokens (Source 1). This goes beyond simply increasing the character count. By remembering a vast amount of information at once, it means its capabilities as a ‘truly working AI (Agentic AI)’ have dramatically improved—such as understanding complex programming code in its entirety and gaining insights that permeate hundreds of pages of financial documents (Source 3, Source 9).
Simply put: The ‘Committee of Experts’ approach
Step5Preview uses a clever method called ‘Mixture-of-Experts (MoE)’ (Source 1).
To use a simple analogy, imagine a school that doesn’t just have one genius student who has to be good at all subjects, but instead has numerous teachers standing by, such as math experts, English experts, and science experts. When a question is asked, not all the teachers rush in; only the math teacher activates to provide the answer for a math question.
Although Step5Preview has a massive total of 600 billion parameters (units of AI knowledge), it selectively uses only 27 billion of them when answering a question (Source 5, Source 9). Thanks to this, it can maintain the vast knowledge of the entire model while capturing the two rabbits of speed and efficiency simultaneously (Source 11).
The AI that has come to our side
Currently, Step5Preview is available for immediate use via API and has the ability to analyze not only text but also video data (Source 5, Source 9).
Industry assessment is also very positive. Analysis prevails that it recorded an intelligence quotient of about 44 points on certain benchmarks, showing top-tier performance even among open-weight models (models where internal information is disclosed) currently on the market (Source 14). It shows unique strengths, especially in precise and professional tasks like software engineering or finance (Source 3). Current usage costs are set at approximately $2.70 per 1 million output tokens (Source 15).
What can we expect?
The biggest event that many developers are paying attention to is October 15. The developer, StepFun, has promised to fully release the model’s weights to the public (Source 6, Source 9). This means anyone can run this powerful model directly on their own server. It is a very interesting point to watch how our daily office environments will change when AI models with dramatically increased memory become popularized.
MindTickleBytes’ AI Reporter Perspective
The ability to memorize vast amounts of data at once will be the key to AI evolving from a simple chatbot into a practical professional assistant. The pace of technological advancement is frightening, but in the end, what matters to us is how we utilize this expanded memory for valuable work.
References
-
[Step5Preview- API Pricing & Providers OpenRouter](https://openrouter.ai/stepfun/step-5-preview) -
[StepFun: Step5Preview· Models · Pi A terminal-based coding agent](https://pi.dev/models/openrouter/stepfun-step-5-preview) - Step5Preview- StepFun Documentation
- Step5Preview API Integration Guide: StepFun’s 600B Agentic…
-
[NVIDIA Nemotron 3 Ultra vs Step5Preview AI Choice Engine](https://aichoiceengine.com/ai-models/nemotron-3-ultra-vs-step-5-preview) - Step 5 Preview: Advancing the Pareto Frontier - stepfun.com
- StepFun shares Step 5 Preview benchmarks, demos… · AGI Hunt
- StepFun Step 5 Preview Technical Analysis — 600B MoE …
- pandaily.com/stepfun-step-5-preview-600b-moe-1m-context.data
- StepFun’s Step5Preview API ships; open weights promised October…
- 1 million token context window
- Can only process text
- Released as a free model
- A structure that always uses all parameters
- A structure that only activates necessary expert parameters
- A technology unrelated to the human brain structure
- Already released
- October 15, 2026
- December 31, 2026