Bottom line up front
Step 5 Preview is a sparse mixture-of-experts model with about 600 billion total parameters and roughly 27 billion active per token. StepFun documents a 1-million-token context window, text, image, and video input with text output, up to 64,000 completion tokens, configurable low/medium/high reasoning effort, and tool calling with structured output. StepFun announced that open weights would be released on 15 October 2026, but its announcement named no licence, and third-party indexes still list the model as proprietary — so the weights are a promise rather than a fact until the licence text is published.
The published specification
StepFun's model documentation lists the identifier as step-5-preview with a 1M-token context window, text/image/video input and text output, and low, medium, or high reasoning effort. Third-party indexes record a 64,000-token completion ceiling, which is well below the context window — the distinction matters, because a large input window does not imply a large output.
The architecture StepFun describes is a sparse mixture-of-experts design: about 600 billion parameters in total with roughly 27 billion active per token. Third-party reporting also records a 92-layer 'narrow-deep' Transformer stack, though that figure comes from StepFun's launch material as reported by others rather than from the model card itself. Architecture details are worth recording with their source, because they change which runtimes can serve the model.
- Record the context window and the completion limit separately.
- Note which figures come from the model card and which from third-party reporting.
- Check whether your runtime supports the architecture before planning self-hosting.
Open weights are a promise until the licence is published
StepFun's announcement states that the model will be released with open weights on 15 October 2026, which is the date this article was written. At the time of writing, the announcement named no licence, and one third-party index still lists the model as proprietary. A stated intention to publish weights is not the same as published weights with a licence.
This distinction is the whole difference between two very different decisions. If the weights land under a permissive licence, self-hosting becomes an option with a hardware question attached. If they do not, or land under restrictive terms, the model remains a hosted API dependency. Either way, the check is specific: read the licence that ships with the checkpoint, not the announcement that promises one.
- Verify whether the weights were actually published, and on what date.
- Read the licence that ships with the checkpoint before planning around it.
- Record the hosted endpoint's terms separately from the weights' licence.
Reasoning effort is a cost and latency dial
StepFun documents low, medium, and high reasoning effort. That setting changes how much deliberation the model does per response, which moves cost and latency alongside depth. One third-party analysis recorded the model generating far more output tokens than the median to run its evaluation index, and described it as verbose — a reminder that reasoning effort and output length are where a cheap-looking model becomes expensive.
Treat the effort level as part of your measured conditions rather than a fixed default. Running the same task at low and high effort on the same model is often more informative than comparing two models at one setting.
- Record the reasoning-effort setting alongside every measured result.
- Measure output tokens, not just input, when estimating cost.
- Test whether a lower effort setting meets your quality bar before paying for the higher one.
Free access is a gateway subsidy
Step 5 Preview is reachable at zero cost on OpenCode Zen. The commercial listing elsewhere records roughly one dollar per million input tokens and two dollars seventy per million output tokens, which is the clearest available indication of the model's actual serving cost. A free endpoint on a gateway is absorbing that cost, with the usual conditions: rate limits, narrower concurrency, and no continuity guarantee.
For evaluation and prototyping that is genuinely useful. For anything that must keep running, design the paid fallback first, or plan the self-hosted path the licence may make possible once the weights actually land.