Jarvis Nelson Osorio: How to Choose Between Small and Large AI Models
Jarvis Nelson Osorio"Bigger is better" is a slogan, not a design rule. I am Jarvis Nelson Osorio, known as Jarvis Osorio: a technology entrepreneur and innovator in Chile who works on AI systems and DAO design, and who also writes and makes art. When founders ask me whether they should "just use the biggest model," I answer with a quieter question: what is the job, and what does a wrong answer cost?
The choice is a portfolio, not a religion
Large models earn their reputation on hard reasoning, messy language, and tasks that still surprise us. Small models earn theirs on latency, cost, privacy, and the ability to run closer to the user — on a laptop, a team server, or a controlled environment. The interesting products I see are not choosing one church. They are building a portfolio.
A useful way to decide:
- If the output must appear in under a blink, a giant remote call is often the wrong default. Micro-decisions — ranking, routing, autocomplete, "is this safe to send?" — favor small, local, or specialized models.
- If the output must be deeply considered, a larger model with sources and a human review step may be worth the wait and the bill.
- If the data should not leave the building, local or tightly contracted models beat a general public endpoint, no matter how eloquent.
- If the task is extraction or classification, a smaller specialist often beats a conversational giant that wants to write an essay about the field you asked it to copy.
I have written about local high-speed agents and about system-one style micro-decisions for this reason. Speed is not a vanity metric. It changes what a product can feel like, and it changes whether a team in Santiago is waiting on a queue in another hemisphere.
A decision card you can actually fill
Before you buy another subscription, fill a card for the feature, not for the company:
- User-visible latency you can tolerate.
- Cost per day at your realistic volume, not at a demo volume.
- Whether prompts or documents contain personal, student, or financial data.
- Languages you must support well — Spanish, Portuguese and English do not always arrive with equal care.
- Whether the model may act (send, delete, pay) or only draft.
- How you will test a substitute if the provider changes terms tomorrow.
If you cannot fill the card, you are not choosing a model. You are hoping.
Hybrids that stay honest
A pattern that keeps paying for itself:
- A small model sits close to the user and handles the fast path.
- A larger model is called when the small model says "I am unsure" or when the task is inherently heavy.
- A checker — rules, tests, or a person — watches the actions that cannot be undone.
- You measure the mix: how often you escalated, how often the small model was enough, how often the large model still needed a correction.
Honesty matters here. Publishing a benchmark that only shows the large model's best paragraph is a marketing exercise. Shipping the hybrid you actually run is engineering.
What I refuse to trade away
There are moments when I will not pick the cheapest small model: anything that talks with children as if it were a companion, anything that gives medical or legal-sounding advice, anything that moves money. Those are judgment surfaces. The Compliant Mind exists because many systems are trained to be agreeable rather than careful. A tiny model that never refuses can be as harmful as a giant one that never refuses. Size is not virtue.
There are also moments when I will not pick the famous giant: tight loops in an editor, on-device helpers, batches you can afford only if the tokens stay inexpensive, prototypes that should run on a developer's machine in Chile without a round trip for every keystroke.
A note on taste
Choosing models is becoming part of product taste, the way typography and permissions already are. The founder's job is shifting from "who has the highest public score" to "which combination respects our users, our latency, and our values." That is curation. I will say more about that shift in other essays; the short version is that you cannot delegate taste to a leaderboard.
Explore more work at jarvisosorio.com and the article hub Jarvis Nelson Osorio — articles, books, videos and profiles. For the human-judgment side of this choice, read The Compliant Mind.
Use the large model when the problem is large. Use the small one when the moment is small and frequent. Keep a way to switch. That is the whole craft, and it is optimistic: we finally have enough good options to choose with intention.