A proprietary model is controlled by its provider: access to the base model, weights, implementation, or execution interface remains restricted. Fine-tuning uses provider tools, incurs per-token training costs, and produces a hosted artifact rather than an exportable model. section 1.1.2
Evaluate proprietary models through their operational trade-offs:
- Treat fine-tuning as provider-dependent: the adapted model remains tied to the provider's base model. section 1.1.2
- Account for recurring cost: training processes tokens through very large models, and inference remains more expensive after adaptation. section 1.1.2
- Plan for lifecycle changes: pinning a version improves stability, but upgrades or retirement can require repeating fine-tuning and evaluation. section 1.1.2
Closed-model fine-tuning
Fine-tuning Gemini or GPT through provider tools adapts behavior without exporting the base weights; switching providers can therefore require rebuilding the adaptation. section 1.1.2
