StepFun Step 5 Preview: What’s Available Now—and What Isn’t

StepFun announced Step 5 Preview on September 20, 2026, as a model for long-running agent work. The distinction to keep in mind is availability: the hosted API is available, while the BF16 open-weights release is scheduled for October 15. A release date is not the same as a downloadable checkpoint or a confirmed license. 3615

What Step 5 Preview offers

StepFun describes the model as a sparse mixture-of-experts (MoE) system with 600 billion total parameters and about 27 billion active per token. That active count describes how many parameters are engaged for a token; it does not mean the full model is a 27-billion-parameter checkpoint. StepFun also lists a one-million-token context window and vision input. 15

Several third-party spec reports describe a narrow, deep design with 92 Transformer layers. One third-party listing gives a different layer count, so treat 92 as a reported specification rather than a point independently settled by the materials cited here. 1619

The model is positioned for software engineering and professional knowledge work, with StepFun highlighting finance. Third-party model descriptions also list long-horizon tool use, research and broader coding tasks. 215

Text and image input are supported in StepFun’s own model description. Third-party listings also report video input, but that capability is not stated in the cited StepFun announcement. 2715

API access and planned weights are different

Step 5 Preview can be used through StepFun’s hosted API. Reported list pricing is $1.00 per million input tokens and $2.70 per million output tokens; some listings also report a $0.05-per-million cache-hit input price. Check the provider’s current pricing before estimating costs, since usage and cache treatment can affect a bill. 367

StepFun has said it plans to release BF16 weights on October 15, 2026. That remains a planned release, not proof that the model can currently be downloaded or run independently. At launch, third-party checks described the relevant Hugging Face repository as lacking model weights; the license and self-hosting terms were also not established in the cited reports. 348

Reported results need context

A launch report relays a 67.7 result on DeepSWE v1.1, while third-party trackers list an Artificial Analysis Intelligence Index score in the range of roughly 43.7 to 44. These are different measures and should not be compared as if they were the same benchmark. The DeepSWE figure is reported by sources covering StepFun’s claims; the index figure comes from third-party listings. Neither number alone establishes that Step 5 Preview is better than another model across tasks. 2623

StepFun also presents the model as offering comparable intelligence at lower task cost. That is a vendor claim; API prices and benchmark scores by themselves do not demonstrate an overall cost or capability advantage in a particular workflow. 1516

What is still unclear about the technical details

Some third-party listings describe attention optimizations such as sparse grouped-query attention and block-wise token merging, while other listings differ on architecture details. The sources cited here do not establish those mechanisms as confirmed StepFun specifications. One secondary report mentions long-horizon reinforcement learning, but the available official description does not provide enough detail to verify the training approach. 6181920

For now, the clearest practical takeaway is straightforward: Step 5 Preview’s API is the available way to try the model; the promised BF16 weights, their license and the exact terms for self-hosting should be checked again if and when StepFun publishes them.