# StepFun Launches Flagship Model Step 5 Preview

- Category: Flash (快讯)
- Published: 2026-09-20T11:56:00.000Z
- Source date: 2026-09-20
- HTML: https://www.88z.ai/en/news/flash-20260920-d7859b2c66d53e0c
- Markdown: https://www.88z.ai/en/news/flash-20260920-d7859b2c66d53e0c.md
- Other language: https://www.88z.ai/news/flash-20260920-d7859b2c66d53e0c.md

StepFun has released its new flagship base model, Step 5 Preview, featuring a sparse MoE architecture, million-token context support, and native multimodality. It ranks among the top three global open-source models in intelligence index, offers significantly lower costs per task than competitors, and excels in long-duration Agent tasks and various professional fields.

## At a glance

- Step 5 Preview ranks top three among global open-source models in intelligence index, with costs at one-eighth of competitors
- Supports million-token context and native multimodality, capable of executing Agent tasks for over 22 hours continuously

## The update

StepFun has officially launched Step 5 Preview, its new flagship base model. This model utilizes a sparse MoE architecture, features native multimodal capabilities, and supports a context window of up to one million tokens. In performance evaluations, it achieved an intelligence index of 44, ranking among the top three global open-source models, just behind GPT-6 Astra and Claude Opus 5. Notably, the cost per task for this model is only one-eighth that of these two competitors, demonstrating high cost-effectiveness. Furthermore, Step 5 Preview can continuously execute Agent tasks for over 22 hours, performing excellently in complex scenarios such as AI programming, financial analysis, and hardware debugging, which highlights its strong capabilities in long-range task processing and professional applications.

## Sources

- [阶跃StepFun](https://mp.weixin.qq.com/s/WLnBfojjQHcuiqI6wteQUw) (2026-09-20)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.