Open weights or proprietary service: check your rights first
Distinguish open weights, open-source AI and proprietary services, and compare rights, deployment control, operations and task quality.
Choose an approach
For self-hosting, verify weight availability and permission for the intended use. For managed operation, compare hosted services. Open weights are not automatically fully open-source or unrestricted commercially; proprietary APIs and data terms also require specific checks.
Weights and open-source rights differ
Open weights provide parameter files, but licenses may restrict use or redistribution. The OSI definition addresses freedom to use, study, modify and share, plus information needed for modification. A download button alone does not establish compliance.
A hosted API delegates inference to a provider. It can serve proprietary or open-weight models. Weight openness and API access are therefore compatible; assess the license separately from the execution model.
Compare outputs on a fixed task
Use actual Chinese classification examples with reference labels and an uncertainty rule. Keep inputs and review criteria identical and record versions. Successful loading or an API response is only the start; inspect long-input behavior and label accuracy.
Specify the needed controls: offline execution, version pinning or weight adaptation. For formatting alone, try instructions and field constraints first. Running an inference service offers no automatic benefit without a relevant requirement.
Evaluate the actual license and version
Keep the license associated with the chosen weight version and record dependencies and hardware. For hosted services, check regions, version policies and data handling. This article does not assess commercial eligibility or equate public downloading with unrestricted use.
Include regression checks and recovery in maintenance costs. Retain a working version and replay the same cases after updates; keep replayable inputs for hosted services too. Choose on observed task quality and maintenance, not an asserted universal ranking.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.