What is the difference between open-weight and closed models?
Open-weight models make their learned parameters available, while closed models generally provide access through a service. The distinction concerns access to the model, rather than a guarantee about quality or price.
Weights influence how a model turns an input into an output. The Open Source Initiative’s definition of open-source AI also covers code, training-data information and freedoms to use, modify and share. Downloadable weights alone do not establish that a system meets that definition. The OSI definition
Licences differ between models. OpenAI’s gpt-oss-20b card lists Apache 2.0 and documents fine-tuning. Those facts concern that model, rather than every model described as open. The model card
What business advantages can open weights offer?
Open weights can give a business more choice over where a model runs and how it is adapted. Those options may matter when existing services cannot accommodate a deployment requirement or a specialised recurring task.
Access also brings operating responsibilities. A business running a model itself has equipment, maintenance and monitoring to account for. A provider hosting an open-weight model may take on some of that work under its service agreement.
The commercial benefit therefore depends on the complete arrangement. Greater control has limited value if the organisation lacks the people or reason to exercise it. A lower processing rate may be offset by integration or support costs.
What business advantages can a closed service offer?
A closed service gives a business access to a provider-operated model without requiring it to run the underlying model infrastructure. Capability, support and existing integrations may make that arrangement suitable for particular work.
Data treatment depends on the product and agreement. OpenAI states that business and API inputs and outputs are not used for model training by default. That commitment does not describe every consumer account or every AI provider. OpenAI’s business-data commitments
A service relationship can also create dependencies on the provider’s availability, terms and product changes. The ability to move records, instructions and connected workflows elsewhere varies with the system.
How can the two options be compared fairly?
A meaningful comparison covers the same work, acceptance standard and expected volume. Model usage is one component of cost alongside integration, review and ongoing operation.
Performance on a narrow public benchmark does not establish suitability for a company’s marketing decisions. Relevant evidence includes factual accuracy, handling of uncertainty and the amount of correction required before work can be used.
Openness and hosting remain separate questions. A cloud service can run an open-weight model, and owning the weights does not require local operation. Local AI and frontier services explains that distinction.
Legal note: This answer provides general information, not legal advice. Seek advice from qualified legal counsel for your circumstances.