Choose a Model

Identify the requirement

Determine the input type and desired output. An available model is not necessarily suitable for every task.

Check the resources

Review the model’s own requirements and the deployment’s capacity. Larger models and longer contexts can increase resource demand.

Verify delivery

For image understanding, confirm both vision support and delivery of the actual image. A filename in a prompt does not establish image access.

Test a representative task

Use a short, checkable example after changing the model. If it fails, inspect model-service status and configuration before treating the result as a quality problem.

Requested model and active model

The Model selector expresses your requested model for future replies. Agent or capability routing can select a different active model. Check the model attribution on the actual reply when comparing results. Changing the selector does not rewrite earlier messages or guarantee that a particular input type is supported.

Ask Friday to guide you

In an authorized FridayLocalAI conversation, try:

Explain the requested Model selector and the active model attribution in FridayLocalAI. Help me design one small test for [text/document/image task] and tell me what to check in the reply. Do not claim a model was used without attribution evidence.

Tell Friday what you can see and which step you have reached. If the instructions do not match your screen, say so; ask for the relevant Help topic or contact your operator. Friday cannot be assumed to see your screen or know your current settings.

Start with the input and result

Choosing a model begins with the task you want to complete. Are you asking for an explanation, editing a paragraph, analyzing readable document text, interpreting an image, or using a separate generation workflow? These are different requirements. A model that is useful for one kind of input may not support another, and a visible option does not establish that every associated capability is configured. Describe the task before choosing based on a name or an advertised size. Identify the input you will provide and how you will check the answer. If the task depends on a source, confirm that the application can supply the relevant contents. A model-selection decision cannot compensate for a document that never became available to the response process.

Read requested and active attribution separately

The Model selector represents the requested choice for future replies. FridayLocalAI’s Agent and capability routing may influence which model actually produces a response. The active attribution shown with the conversation or reply is therefore important evidence. Check it when a result is unexpected or when you are comparing models. Earlier replies retain their own history; changing a selector does not retroactively regenerate them. If you want a comparison, make a new bounded request and record its actual attribution. Keep the prompt, source material, and review criteria comparable. Otherwise, differences in the task or context can be mistaken for differences in the model, leading to a conclusion that the evidence does not support.

Consider resources without assuming a universal rule

A local LLM uses resources available to its deployment. Model size, context, simultaneous activity, and the selected execution path can affect suitability. The application includes model-management and safety-related controls, but the options available to you depend on the installation and your role. Ask the operator about a supported choice when the requirements are unclear. Do not assume the largest available model will always give the most useful answer. A smaller option may be appropriate for a straightforward task, while a more demanding task may require a different capability or better source context. The right comparison includes accuracy, usability, responsiveness, and the amount of correction needed. Hardware specifications alone do not settle the quality of a particular result.

Run a small comparison you can evaluate

Choose a task with an answer or review standard you understand. For example, supply a short paragraph with known facts and ask for a structured summary without added details. Add a case with an intentionally missing value and check whether the response leaves it unknown. If document work matters, include a narrow source question after verifying attachment access. Record the active model, the prompt, and the outcome for each run. Avoid judging solely by length or confidence. Note whether the answer followed constraints, preserved facts, and acknowledged uncertainty. If a result changes, repeat only as needed to understand the difference. A small, documented comparison is more useful than switching randomly among models and remembering only the most impressive response.

Recognize when the problem is elsewhere

A failed answer can result from missing source content, an unavailable service, an unsuitable input type, or a misunderstood request. It is not always a model-quality problem. Before changing the model, ask what evidence points to the failing stage. Check whether a short text prompt works and whether the issue follows a particular attachment or workflow. If the response claims to see an image but only a filename was supplied, the important question is whether the actual image reached a supported path. If a document summary lacks the requested section, check source access and context. Changing the selected model without addressing those conditions may produce a different-looking answer while leaving the underlying problem unresolved.

Use Friday to plan the decision

Ask Friday to explain the relevant controls and help design a comparison using the options you can see. Supply the model labels and your intended task without expecting it to know the machine’s resources automatically. Ask it to identify what remains unknown and what should be confirmed with the operator. If you do not have permission to manage models, use the supported choices made available to your role. Do not download or alter system components based on an improvised conversational instruction. FridayLocalAI early access follows an operator-supported deployment path, and the public website does not provide a universal installation recipe. A good model choice is one you can explain in terms of the task, available capability, observed result, and practical limits of the environment.