January 21st.Step StarTwo new models in the Step-2 series of language models were launched yesterday -- Step-2 mini, a smaller participant size and more cost-effective model, and Step Literature Master, a model specifically for the content creation field.
1AI learned from the official introduction that Step-2 mini and Trillion ParametersLarge Model Compared with Step-2, it retains its modeling performance above 80% with a parameter count around 3%.
Meanwhile, Step-2 mini Faster generation speeds and excellent value for moneyThe average initial word latency of Step-2 mini is only 0.17 seconds with 4000 tokens. In the case of inputting 4000 tokens, the average first-word latency of Step-2 mini is only 0.17 seconds. At present, you can already call the API interface of Step-2 mini on the open platform of Step Star. Input 1 yuan/million tokens; Output 2 yuan/million tokens.
Step-2 mini adopts a new attention mechanism architecture independently developed by Step-Star - MFA (Multi-matrix Factorization Attention) and its variant MFA-Key-Reuse, which saves nearly 94% KV cache overhead and significantly reduces inference cost compared with the commonly used MHA (Multi-Head Attention) architecture. Compared with the commonly used MHA (Multi-Head Attention) architecture, it saves nearly 94% of KV cache overhead, has faster inference speed and significantly reduces inference cost.
According to the official introduction, Step-2 Literary Master Edition is a model developed specifically for the creation of textual content, following Step-2's knowledge base, the ability to control the text of powerful details.Featuring more robust content creation capabilitiesStep-2 Literary Masters Edition seeks to solve the problem of over-alignment of language models in the market, which leads to "false and empty" content and a lack of novelty and true feelings.