Model weights are the learned numerical parameters used with a model’s architecture and inference software to produce outputs. Access to those parameters removes one dependency on a supplier; it does not transfer the complete service built around them.
Portability is the ability to move that service to another environment while preserving its required functions. Replacing its model with a different model is a substitution as well as a migration, so it creates an additional question about behaviour.
Access, rights and hosting
Downloadable weights establish access to an artifact. The applicable terms determine which uses and modifications are permitted. The Open Source Initiative’s Open Source AI Definition 1.0 goes further than weight availability: it specifies freedoms and a preferred form for modification that includes code, data information and parameters. The definition does not make every downloadable model an open-source system. Open Source AI Definition 1.0.
Hosting is a separate choice. A third party can operate a service using downloadable weights. An institution can operate those weights itself if it has the required rights, software and infrastructure. The location of inference and the services receiving requests determine operational custody of data.
The dependency is the deployed service
A deployed service combines the model with a tokenizer, configuration, prompts, retrieval, tools and application logic. Exporting only the weights leaves those other parts behind. A provider API can also supply functions the replacement runtime does not implement.
Each dependency therefore needs either a usable replacement or a preserved artifact. A migration record identifies the versions and configurations that produced the old behaviour. Provenance supports checking which artifact was obtained and how it was built; it does not demonstrate that a replacement service gives acceptable answers.
Moving the same model narrows the change, but does not prove the entire service is unchanged. Tool permissions, prompt assembly and retrieval filters still depend on the surrounding application. A migration that preserves model output while widening document access has changed the service’s control boundary.
Retrieval indexes carry model assumptions
An embedding index stores document vectors produced by an embedding model. A query vector must be meaningfully comparable with those stored vectors. Equal vector length establishes only that the arrays have the same number of coordinates.
Take a document vector [1, 0] and a matching query represented as [1, 0]. Their cosine similarity is 1. Now swap both coordinate axes: the document and query both become [0, 1], and similarity remains 1. Swap only the query while leaving the stored document unchanged, and similarity becomes 0.
That construction shows why preserving dimensions is insufficient. A change to the representation requires a compatible document index or a validated alignment. Changing the text-generation model while retaining the same embedding process is a different migration and does not, by itself, require new document vectors.
A portability test ends with service behaviour
A usable exit test runs representative cases in the replacement environment. It checks retrieved evidence, access boundaries, tool actions, response quality and resource requirements against the service’s acceptance criteria. A successful download or startup establishes neither those results nor the cost of sustaining them.
Weight access makes independent operation possible under the right conditions. Demonstrated portability means those conditions have been assembled and the required service still works.