Language and vision models come in two broad forms: proprietary models you access through an API, and open-weight models whose weights you can download and run.
Proprietary API Models
Pros: usually the most capable; no infrastructure to manage; fast to start; continually improved. Cons: data leaves your environment; costs scale with use; the provider controls updates, pricing and availability; limited customisation.
Open-Weight Models
Pros: run on your own infrastructure, keeping data in-house; predictable costs at high volume; can be fine-tuned freely (subject to licence); no dependence on one provider. Cons: you manage hardware, scaling, security and updates; the most capable open models need substantial GPU resources; performance may trail the best proprietary models for some tasks.
"Open" Isn't Always Open Source
Many models release weights under custom licences that restrict commercial use, user numbers or certain applications. Read the licence and the model card.
Choosing
- Sensitive data or strict residency rules: favour self-hosted open-weight models, or providers with suitable data commitments.
- Fast prototyping and top quality: start with an API.
- High, steady volume: self-hosting may be cheaper; do the maths including engineering time.
A Hybrid Approach
Many organisations prototype with APIs, then move specific high-volume or sensitive workloads to open-weight models once requirements are clear.