On the eve of the 2025 World Artificial Intelligence Conference (WAIC2025), Stepwise Star announced its new generation of foundational large model - Step3, in Shanghai. This model not only achieves breakthroughs in intelligence but also considers reasoning efficiency, aiming to provide enterprises and developers with the best application solutions. Step3 is planned to be officially open-sourced on July 31, marking Stepwise Star's contribution of a powerful multimodal reasoning model to the global open-source community.
To promote the practical application of large models, Stepwise Star has also joined forces with several leading domestic chip and platform manufacturers to establish the "Modular Core Ecosystem Innovation Alliance". The goal of this alliance is to accelerate the application of large models in various industries through innovative cooperation between models and the chip industry chain. In addition, Stepwise Star has reached a strategic cooperation with Shanghai State-owned Capital Investment Co., Ltd., aiming to strengthen the collaboration of capital and ecosystem business.
The founder and CEO of Stepwise Star, Jiang Daxin, said that the rapid iteration from Step1 to Step2 has made the team more focused on customers' real needs when considering the most suitable models for applications. The new Step3 model is defined as the best application model for the reasoning era, featuring strong intelligence, low cost, open source, and multimodal capabilities, all four dimensions are indispensable.
Step3 is Stepwise Star's first fully-sized native multimodal reasoning model, with a parameter count of 321B, including 38B activated parameters. It performs well in visual perception and complex reasoning, capable of handling cross-domain knowledge understanding and visual information analysis. For example, it can accurately read the content of reflective menus, and even automatically calculate the cost-sharing in WeChat chat records. Step3 has achieved SOTA results in open-source multimodal reasoning models on multiple rankings.
To improve the reasoning efficiency of the model, Stepwise Star has innovated in the architecture design, enabling Step3 to perform well in reasoning on different hardware platforms. According to tests, the reasoning efficiency of Step3 on domestic chips can reach up to 300% of DeepSeek-R1, and in distributed reasoning with NVIDIA Hopper architecture chips, the throughput has increased by more than 70%.
Stepwise Star has established a "1+N" model matrix, where "1" refers to Step3, and "N" refers to multiple multimodal models covering areas such as speech, visual understanding, and image generation. Through cooperation with nearly 10 chip and infrastructure manufacturers, Stepwise Star is committed to improving the adaptability and computing efficiency of large models, thus promoting the implementation of applications.
Key Points:
🌟 The new generation of foundational large model Step3 will be open-sourced on July 31, marking a major advancement in multimodal reasoning models.
🤝 Stepwise Star has jointly established the "Modular Core Ecosystem Innovation Alliance" with multiple chip and platform manufacturers, accelerating the practical application of large models.
📈 Step3 demonstrates excellent performance in reasoning efficiency, achieving efficient operation on various hardware platforms.