The Terms Are Public. Now They Matter.
The StepFun Model License 1.0 arrived on the HuggingFace repo October 15, 2026 – the same day the weights landed after their brief absence. The gap is closed. The slip is resolved. What is left is the document itself, and it is not what “open weights” usually means. Apache-style in structure, yes. OSI-approved? No. Three additional conditions sit between the download and the deploy, and each one gates a different kind of power.
Three Conditions, Three Gates
The license layers its restrictions by who you are and how big you are getting.
Attribution at scale (Section 2b). If any commercial product or service using the model exceeds 100 million active users or generates more than $20 million in monthly revenue, the operator must prominently display “Powered by StepFun AI” on the user interface and in documentation. For most deployers, this is cosmetic – a branding tax on success that costs nothing until you are already winning.
The hardware OEM gate (Section 2c-i). This is the structural one. Device manufacturers – phones, vehicles, smart glasses, wearables, robots, smart-home devices – must enter a separate commercial license with StepFun before making Step-5 capabilities available as a device feature. The clause covers on-device deployment, remote deployment, infrastructure-supplied delivery, and whether the feature arrives at shipment or later by activation, download, or update. There is no self-serve path. If you build hardware and want Step-5 inside it, you negotiate first.
The MaaS gate (Section 2c-ii). Model-as-a-Service operators with combined revenue exceeding $20 million over any consecutive 12-month period must also enter a separate commercial license. The license defines MaaS as “giving a third party inference or fine-tuning access to the Model or a Derivative Work, including through an API, where the third party has meaningful control over inputs, parameters, or training data.” Critically, “merely relaying requests” to another provider does not constitute MaaS. That carve-out protects pure intermediaries but not providers offering parameter control or fine-tuning – and the $20 million threshold is low enough that any significant inference business is in scope.
There is a safe harbor. Section 2(c) does not apply to internal use that does not make the model available to any third party. If Step-5 stays inside your infrastructure and your data, the OEM and MaaS gates do not bind.
The Restrictions Follow the Model
The Derivative Work definition in Section 1 extends the reach. Fine-tuned, distilled, merged, quantized, pruned, and continually pre-trained versions all count as Derivative Works and inherit the same license terms. A model trained to a material extent on Step-5 outputs is deemed to reproduce it. But a separate application that merely calls the model through an API is not a Derivative Work – the license draws the line at modification, not consumption.
This is where “open weights” stops being a shorthand for “deployable.” The weights are downloadable. The deployment categories that matter most – hardware integration and inference-as-a-service at scale – are gated behind commercial negotiation.
No Other Lab Does This
The comparison is stark. Kimi K3 ships under a modified MIT license. Meta’s Muse Glimmer 30B uses Apache 2.0. GLM-5.3 Flash uses MIT. None of these carry OEM or MaaS gates. None restrict device embedding. None require a separate commercial license for inference providers above a revenue threshold. StepFun has built something different: an open-weight release that preserves the download while controlling the deployment.
What Deployers Must Do
Infrastructure planners need to audit their deployment architecture against these three gates before committing to Step-5. If you are building a hardware product, the OEM gate is binary: negotiate or look elsewhere. If you are running an inference service above $20 million in combined annual revenue, the MaaS gate applies. If both, both.
The internal use carve-out remains the cleanest path. Organizations that can keep Step-5 within their own infrastructure – no third-party access, no device embedding, no external inference service – avoid the commercial gates entirely. For everyone else, the license turns model selection into a legal and financial decision.
Where This Leaves the Calendar
The calendar taxonomy now has its terms exhibit. The vendor-promise class produced both outcomes within a single day: the weights missed their date, then landed. But the license that arrived with them is the part that changes planning. The gate is no longer at the download button. It is at the point of deployment, and it has three conditions. Whether “open weights” means “deployable” now depends on who is asking and how they intend to use it.