Step 5 Preview reaches the API, with open weights due October 15
StepFun made Step 5 Preview, a model with 600 billion total parameters and 27 billion active per token, available through its API. The company says open weights for the software engineering and extended tool-use model are due October 15, so today's delivery remains a hosted service rather than a local deployment package.
Artificial Intelligence··Midday
27 billion of 600 billion activate per token
Chinese AI company StepFun introduced its new flagship, Step 5 Preview, with a sparse mixture-of-experts architecture. The model has 600 billion parameters in total and activates 27 billion of them for each token. The figures describe the complete weight set and the subset used during one processing step. The company also announced a 1 million-token context window together with text and image input. The design uses expert subsets chosen by a router instead of running every parameter for each token; the announcement did not specify how those experts would be placed across hardware.[1], [2]
The model targets software and tool tasks
StepFun positions Step 5 Preview for AI coding, software engineering, professional knowledge work and finance. The announcement emphasizes workflows that make multiple tool calls across a long context and continue a task through several steps. The open API lets developers use the model through a service hosted by the company. Published performance comparisons remain StepFun's own measurements. StepFun said its StepCodeBench measurement covered 553 repositories, 9 task types and 33 programming languages. It also claimed a score of 44 on the Artificial Analysis Intelligence Index and a place among the top three open models; these are company measurements rather than independent tests.[1], [2]
Open weights carry an October 15 date
The API became available on September 20, but the weight files were not released at the same time. StepFun says the open weights are due October 15, creating a stated interval between hosted access and the point when users can deploy the model on their own hardware. Weight precision, compression formats and device-placement options have not been announced, so local memory requirements are not yet specified. The company already offers a 1 million-token context and image input through the API. Local deployment options remain dependent on publication of the weight package and its technical details.[1], [2]
Related columns
For more information on this topic, you can read the related columns.
