According to CoinMeta, Step Star has released the weights of 5 preview models with the identifier Step. These models feature a deep architecture with 92 layers and 600 billion parameters, and are capable of handling 1 million token contexts. A technology blog revealed that this model introduces sparse and gqa to merge adjacent token indices, reducing the indexing and filtering costs to 1/8 of those of traditional index schemes with token per index. The model utilizes a custom set of stepfun, model, and license versions 1.0, which allows for general commercial use. However, hardware manufacturers such as those in the mobile phone and automotive industries that wish to integrate this model's capabilities into their products must sign a separate licensing agreement.