Shanghai Step Star Intelligent Technology Co., Ltd. released two new members of the Step series of language models on January 20, 2025: Step-2mini and Step-2 Literature Master Edition. This is an important step taken by Stepstar to meet the diverse needs of developers and further expand its artificial intelligence application capabilities following the launch of Step-2, a large trillion-parameter language model last year. With its industry-leading performance, Step-2 has repeatedly won success on international authoritative lists, providing developers with powerful AI application support. The launch of this new member aims to improve the response speed, cost performance and creative capabilities of the model to better serve different creative scenarios.
On January 20, 2025, Shanghai Step Star Intelligent Technology Co., Ltd. announced the launch of two new members of its Step series of language models-Step-2mini and Step-2 Literature Master Edition. This initiative aims to meet the differentiated needs of developers for language models in different creative scenarios and further expand its application capabilities in the field of artificial intelligence.
Following the launch of the trillion-parameter language large model Step-2 last year, Step Star has repeatedly ranked first among domestic large models on international authoritative lists such as LiveBench with its industry-leading model performance. The release of Step-2 provides developers with powerful tools to support multiple AI application capabilities such as plot creation, role playing, and copywriting. However, as application scenarios continue to expand, developers have put forward higher requirements for the response speed, cost-effectiveness, and creative capabilities of language models. To this end, Step Star has launched two new members of Step-2Family to meet diverse market needs.
Step-2mini is a lightweight agile model that retains more than 80% of the model performance of Step-2 with only about 3% of the parameters. It significantly improves the generation speed and cost-effectiveness while maintaining efficient performance. When 4000 tokens are input, the average first word delay of Step-2mini is only 0.17 seconds, showing extremely fast response capabilities. In addition, the model adopts a self-developed new attention mechanism architecture - MFA (Multiple Matrix Factorization Attention) and its variant MFA-Key-Reuse, which saves nearly 94% of KV compared to the traditional multi-head attention architecture. Caching overhead greatly reduces inference costs. Currently, developers can already call Step-2mini’s API interface through the Step Star Open Platform. The input cost is 1 yuan/million tokens and the output cost is 2 yuan/million tokens.
At the same time, the Step-2 Literature Master Edition focuses on content creation and aims to provide more powerful support for creators. This version inherits Step-2’s knowledge reserve and ability to control text details, and further strengthens content creation capabilities. Different from the "fake and empty" content generated by some large models, Step-2 Literature Master Edition can fully understand the creative needs of users and generate works with strict logic, concise language, rich ideas and unique style. Whether it is suspense novels, cool novels or science fiction novels, Step-2 Literature Master Edition can show its outstanding creative strength. For example, the science fiction novels it generates are not only based on cutting-edge scientific knowledge, but also incorporate humanistic philosophy, showing profound themes and fascinating narrative styles.
Currently, the Step-2 Literature Master Edition has been launched on the Yuewen App, and the web version will also be fully launched this week (https://yuewen.cn) for users to experience. Step Star stated that it will continue to launch more new models intensively before the Spring Festival to provide users with more comprehensive and powerful general-purpose large models.
The launch of Step-2mini and Step-2 Literature Master Edition marks Step Star’s continuous innovation and breakthrough in the field of large-scale language models. It also indicates that in the future, AI technology will better serve all walks of life and create more for users. value. Look forward to Step Stars bringing more surprises in the future.