Choosing a model

You don’t need to know model names

There’s a lot of jargon around AI models. Most of it isn’t useful when you’re trying to build a working agent.

The truth is: for most agents, the right choice is to let VDF AI pick for you. The platform knows the catalog, knows what each model does well, and adapts to the kind of work the agent is doing. You focus on the job; the platform focuses on the model.

When you do want to pick — because you’ve tested options or because compliance requires a specific model — the choice comes down to a few clear trade-offs. This page walks through them.

The four trade-offs that matter

Every model choice comes down to four levers. You don’t need to know any model names to think about these.

Fast vs. thorough

The most common trade-off.

Fast models are ideal for:

• Short answers — definitions, classifications, simple summaries.
• High-volume work — running the agent hundreds or thousands of times.
• Real-time interaction — the user is waiting for a reply.

Thorough models are ideal for:

• Long-form drafting — a release note, a brief, a multi-section analysis.
• Nuanced reasoning — a critique, a strategic recommendation, a careful comparison.
• Complex inputs — large source documents, multi-step instructions.

A useful test: read the agent’s instructions out loud. Does the job sound like something you’d ask of a sharp colleague in 30 seconds, or in 30 minutes? If 30 seconds, fast. If 30 minutes, thorough.

Cloud vs. on-premises

Two practical reasons to choose on-premises:

• Data residency. Some organizations require AI to run inside their own environment.
• Sensitive content. Customer data, financial details, regulated information — your organization may prefer this never leaves your infrastructure.

For everything else, cloud is the broader catalog and usually the easier choice. The platform’s privacy posture means cloud is safe for the vast majority of work — see Privacy & Security.

If you’re unsure, ask your workspace admin. They’ll know your organization’s posture on this.

Text-only vs. image-capable

A simple question: will your agent ever look at images?

• Reading screenshots a teammate attached.
• Analyzing charts or diagrams.
• Looking at product imagery.
• Generating descriptions of images.

If yes, pick a model that supports images. If no, a text-only model is usually cheaper and faster.

The library agents that produce images — Image Generator, Chart Generator, HTML Mockup Generator — already use the right model for their work. You don’t have to think about it.

Cost vs. quality

For most one-off and low-volume agents, cost barely matters. The difference between models is fractions of a cent per run.

It starts to matter when:

• The agent runs hundreds of times a day.
• The agent processes large inputs — long documents, many records.
• The agent is part of a customer-facing product where the cost is per customer.

A useful pattern: start with quality, optimize for cost later if needed. Use a high-quality model while you’re tuning the agent. Once the agent is producing reliable results, test whether a cheaper model produces nearly-the-same quality. If yes, switch.

When to let VDF AI pick

The default and the right choice for most agents. VDF AI looks at the agent’s job and picks a model that fits — usually a balanced choice that’s fast enough, thorough enough, and reasonably priced.

The advantage: as new models arrive, your agent benefits automatically. The model that backs your agent six months from now will likely be better than today’s, without you having to do anything.

This is similar to smart routing in Networks, with the same underlying principle: let the platform handle the catalog, you handle the work.

When to pin a specific model

A few situations where pinning is the right call.

Compliance

If your organization requires a specific model for a specific kind of work, pin it. The platform doesn’t second-guess compliance requirements.

A tuned model

If your team has fine-tuned a model on your data (see Fine-tuning datasets), pin it. That’s why you tuned it.

A tested winner

If you’ve genuinely tested several models on the agent’s job and one is clearly better, pin it. But test honestly — most teams find that “clearly better” is rarer than they expect.

A cost ceiling

If the agent runs at very high volume and you’ve decided which cheaper model is good enough, pin it. Cost compounds.

A useful sequence for a new agent

If you’ve never thought about model choice before:

1. Build the agent with VDF AI picking the model. Don’t decide upfront.
2. Use the agent for a week. Pay attention to whether the output feels right.
3. If something’s off, look at the model used. It’s logged for each run.
4. Decide whether to pin or to keep auto. Often the answer is “keep auto” — most teams’ agents work well at the default.

What to do when an agent’s output changes

If you’ve been using an agent and the output starts to feel different — slower, faster, longer, shorter, more or less detailed — the model behind it may have shifted. The platform updates auto-picked models as better ones arrive.

Two checks:

• Is the new output better, worse, or just different? Sometimes “different” is “better” and you weren’t expecting the change.
• Is the change consistent? If yes, that’s the new normal. If no, the model picker may be choosing differently per run — usually because the inputs vary.

You can pin a model anytime to freeze the behavior. Pinning is reversible — you can switch back to auto-pick later.

Where to go next

• Creating your own agent — set the model preference during agent creation.
• Tools and knowledge — model choice matters less when sources are strong.
• Smart model routing — the same concept applied to Networks.
• Chatting with agents — what to expect from a model in conversation.