Term: Open-weight models
~6 min read
Estimated time: ~6 min read — for the in-app brief plus opening the primary source.
What this is
Open-weight models ship the trained weights so you or a host can run them. You gain control over location and customization — and you own more of the operations.
Everyday example
A vendor offers to run a Llama-class model in your cloud so prompts never leave your tenancy. That is an open-weight (or open-source-weight) choice — different from calling ChatGPT’s closed API.
Open-weight models ship the trained weights so you (or a host) can run them — different control and cost than a closed API.
- Llama-style models are the familiar example; ChatGPT-class flagships are typically closed APIs.
- Upside: more control over data location and customization.
- Downside: you own security, updates, and evaluation.
- “Open” is not automatically safer or cheaper once you add people and infrastructure.
Next action: If a vendor says “open source,” ask whether you run the weights, and who patches them.
What changes in how you lead
How decision rights, process, and ownership should change.
- Purchasing treats open-weight as a make-or-buy on operations, not a slogan.
- Risk and IT must own runtime security if the model sits in your environment.
Compare related ideas
Open-weight models vs Closed lab APIs
Labs’ flagship chat products are usually closed APIs. Open-weight models let you (or a host) run the engine. Control and operational burden move toward you.
Open Closed lab APIsDeep dive
Closed APIs (typical ChatGPT / Claude / Gemini / Grok enterprise) keep the weights at the lab.
Open-weight (Llama and peers) can run in your environment or a specialist host.
Ask who patches, who evaluates, and where data sits.
Purchasing and Risk treat this as supplier and runtime risk, not a slogan.