AI model selection and licence review
The right model for each use case, assessed on performance, language coverage, licence terms and origin.
Overview
Lindstead selects AI models on evidence: each candidate is tested against the organisation's own use cases and languages, its licence is checked for what a company of that size may do, and its origin and supply chain are reviewed. The result is a shortlist the organisation can defend to auditors and the board.
Key questions
- Which open-weight models perform well enough for our use cases and languages?
- May a company of our size use this model commercially under its licence?
- Does the model's country of origin create policy or supply-chain risk?
- How far behind the closed frontier is the best model we can run ourselves?
Approach
- Use-case evaluation Candidate models are evaluated on representative tasks from the organisation, in the languages it works in, with results recorded per task.
- Licence review Each licence is read for commercial use, revenue or user thresholds, regional restrictions and redistribution terms. Some well-known licences exclude large companies or EU entities.
- Origin and supply-chain review Developer, jurisdiction, training disclosures and known evaluations of model behaviour are reviewed, so model origin becomes an explicit policy decision.
Deliverables
- Model shortlist Recommended models per use case, with evaluation results and trade-offs.
- Licence review What each licence allows the organisation to do, in plain language.
- Origin assessment The supply-chain and policy considerations for each candidate model.
Frequently asked questions
-
An open-weight model publishes its trained weights so it can run on your own infrastructure. That does not make it open source: the licence can still restrict commercial use, company size or region, and training data is often not disclosed.
-
It depends on the licence of each model and version. Lindstead's Model Index lists the leading models with origin, licence terms for large companies and minimum hardware, updated monthly.
-
Running weights locally removes the data transfer to the developer, but not behaviour trained into the model. Lindstead treats model origin as a board-level policy question and documents the trade-off.
Related
Discuss model selection with Lindstead.
Schedule an introductory meeting with our team.