Several installed models should not leave users guessing which one produced their work. Show the active configuration and explain differences in terms of supported tasks and device requirements.
Keep compatibility a product decision backed by evidence. A file being present or loadable does not establish that it meets the app's output, resource, or source-handling contract.
Separate installed, active, and supported
An installed model may be retained for recovery without being the current default. An active model may support some tasks but not a newly added feature. Record these meanings explicitly instead of using one ready badge for every state.
Maintain a versioned set of model artefacts, runtime and bridge versions, backend requirements, prompt contracts, task limits, and tested hardware. Consult current runtime documentation and the exact model publisher's requirements.
Review a fictional two-model workspace
Suppose a laptop has a smaller supported model and a larger optional model. The user starts a short document draft, switches the selected model, and then opens an older saved result.
The in-flight request should retain its original configuration, while a new task follows the new explicit choice. The saved result should remain readable and identify relevant provenance without requiring the old model to load merely for viewing.
Test incompatible model, missing file, failed update, unsupported task, cancellation during switching, and account or document change. Verify late output cannot be attributed to the wrong model or overwrite a reviewed draft.
Keep the choice understandable and bounded
Describe useful differences such as supported input, resource requirement, or evaluated task scope. Do not label one option best or accurate without the appropriate comparative evidence.
Provide clear cleanup choices and explain when removing a version affects fallback or offline capability. Avoid exposing internal filenames as the only way to identify a configuration.
The tradeoff is flexibility versus evaluation and support complexity. More combinations can accommodate varied hardware but multiply regression cases and customer explanations. A narrow supported window may serve users better than an unverified model catalogue.
Keep support diagnostics tied to the actual model used by the operation, not the latest setting shown in the interface. That distinction matters when users change preferences while work is running.
The ChalkLens portfolio account documents local inference in a teaching workflow. This two-model design is fictional and does not claim a current model selector, compatibility matrix, or benchmark for that product.
Questions about several offline models
Should changing the default alter an active request?
Only through an explicit supported operation. Preserve configuration identity so results and resources remain attributable.
Can old results be opened without their original model?
When the product stores ordinary output independently, yes. Define and verify that lifecycle instead of assuming it.
What should a review inspect first?
Inspect installed versus active state, task compatibility, in-flight switching, result provenance, and cleanup. Bring confusion to a local AI versioning review.