StepFun 227

StepFun is an innovative company focused on the field of artificial intelligence, founded in April 2023. With the mission of "Intelligence Leaps, Multiplying the Possibilities for Everyone," it aims to achieve Artificial General Intelligence (AGI). The company was founded by Jiang Daxin, former Global Vice President of Microsoft. After leaving Microsoft, he led the team to explore the Scaling Law, dedicated to developing and optimizing large model technology.

StepFun's core competitiveness lies in its self-developed super models, including a hundred-billion-parameter language large model, a hundred-billion-parameter multimodal large model, and a preview version of a trillion-parameter MoE language large model. These models perform excellently in logical reasoning, Chinese knowledge, and English knowledge, while covering multiple aspects from personal use to enterprise-level applications. Notably, the Step-1V multimodal large model demonstrates technological innovation and performance leadership on the "Sinan" (OpenCompass) evaluation leaderboard.

StepFun's products include two applications targeting the consumer market: Yuewen and Maopao Ya, both of which are fully open for use.

StepFun Core Features:

• Language Processing Capability: The Step-1 hundred-billion-parameter language large model comprehensively outperforms GPT-3.5 in logical reasoning, Chinese knowledge, English knowledge, mathematics, and code. This indicates that the model has powerful capabilities in understanding and generating text, capable of handling multiple languages and professional knowledge.

• Multimodal Capability: The Step-1V hundred-billion-parameter multimodal large model excels in image understanding, multi-turn instruction following, and mathematical abilities. This means the model can not only process textual information but also understand and generate image-related data, as well as execute complex instructions and perform mathematical calculations.

• Deep Intelligence Exploration: The Step-2 trillion-parameter MoE language large model adopts the MoE architecture and focuses on the exploration of deep intelligence. This indicates that while providing API interfaces for partners to try out, the model also explores how to further enhance the intelligence level of AI through larger model scales.

• Embodied Intelligence Applications: Applying multimodal models to robots or other devices, enabling them to interact with the physical world, autonomously collect data, build world models, and incorporate complex task planning capabilities on this basis. This feature explains that StepFun's models are not limited to processing virtual data but can also directly act on real-world objects and environments, achieving more intelligent application scenarios.

To promote industry development and innovation, StepFun has adopted a "closed-source + open ecosystem" strategy, opening up its model capabilities to the outside world through APIs and other means, encouraging more developers and enterprises to develop applications, and jointly creating application scenarios and industry ecosystems. In addition, the company has collaborated with Jiemian Financial Union to jointly build a large model for the financial industry, promoting industry innovation and upgrading.

StepFun's technology roadmap and product strategy have gained market recognition. The company received investments from multiple top-tier VC institutions in its early stages. The company's success is reflected not only in its technology R&D but also in its comprehensive layout of the four key elements—computing power, data, algorithms, and systems—as well as its unified pursuit of multimodal understanding and generation.

In summary, through its innovation and application in large model technology, StepFun has demonstrated its leadership and technical strength in the field of artificial intelligence, taking a solid step toward the goal of achieving Artificial General Intelligence.