Small vs large AI models, explained with a bicycle courier and a lorry
Large frontier AI models handle the hardest work but cost more and run slower; small models are quick and cheap for focused jobs. Explained with a bicycle courier and a lorry.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
Google described its new Gemini 4 Hydrogen as a fast, cost-efficient model that sits below its two frontier models, and other model launches this week drew the same line between big and small.
In one sentence: AI companies sell models in sizes: large "frontier" models handle the hardest, most open-ended work but cost more and answer more slowly, while small models are quicker and cheaper and do well on simple, focused tasks, so the smart choice is to match the model to the job.
The everyday picture
Picture a delivery company. It has bicycle couriers and big lorries (trucks). Both deliver things. Neither is simply "better".
A courier is fast through busy streets and cheap to send. It is perfect for an envelope or a small parcel. A lorry costs much more to run and takes longer to load, but it can carry a whole house move.
A good dispatcher does not send a lorry to deliver a letter. AI models work the same way. The small model is the courier. The large model is the lorry. Your job is to be the dispatcher.
A dispatcher sends simple jobs to the quick, cheap courier (a small model) and heavy jobs to the lorry (a large frontier model). Both deliver.
How it works, step by step
- A task arrives (a delivery order). It might be "sort this email into a folder" or "plan a migration of our whole billing system".
- Someone decides which model gets it (the dispatcher). This can be a person picking a model in settings, a rule in your software, or a small "router" model. Anthropic's guide to choosing a model says to weigh capability, speed and cost.
- Simple, repeated jobs go to a small model (the courier). Microsoft's overview of small language models says they answer with lower delay and cost less to run, and lists jobs like customer FAQs, short summaries and sentiment checks.
- Hard, open-ended jobs go to a large frontier model (the lorry). "Frontier" means a company's most capable models. Google's Hydrogen model card says its frontier models are built for deep, long reasoning on the hardest problems.
- Often both work together (the courier hands big loads to the lorry). Google says Hydrogen can work as a helper alongside frontier models. Anthropic describes pairing a cheaper model that does most of the work with a frontier model that handles the hard decisions.
Try it · Side by side
Courier or lorry: which model fits each need?
Tap what matters to you (pick one or more) and see which option fits.
Small model (courier)
- Lower: cheap to run for high-volume work.
- Faster, with less delay for waiting customers.
- Can miss the point on broad or complex questions.
- Often does them well enough.
- Quality can drop on long, tangled tasks.
Large model (lorry)
- Higher: more computing power per answer.
- Usually slower, especially on long reasoning.
- Built for deep, multi-step reasoning.
- Works, but you pay for power you do not use.
- Handles bigger, messier jobs more reliably.
Pick what matters to you and the better fit lights up.
Where the picture stops
- A courier can never carry a sofa. A small model can sometimes do a "big" job well enough, and you only find out by testing it on your own tasks.
- Vehicles come in fixed sizes. AI models also have settings, such as how much "thinking" they do, that shift speed and cost within one model. Anthropic says adjusting this is often a better lever than switching models.
- Small models are not always separate builds. Google says Hydrogen was adapted from its larger Carbon model, so the courier and the lorry can come from the same factory.
Why it matters for your business
- Cost. Most business AI work is high-volume and simple. Sending it all to the biggest model is like hiring a lorry for every letter.
- Speed for customers. Small models answer with less delay, which matters for chat, search boxes and anything a customer waits on.
- Quality where it counts. Large models are more accurate on broad, complex questions. Microsoft notes small models can be less accurate on broad queries.
- Less risk of a bad bet. Anthropic offers two starting points: begin with its smallest, cheapest model and upgrade only where it falls short, or begin with a frontier model and step down once results are good. Either way, it says a good set of test questions is the most important step.
An example
Imagine a 30-person travel agency that wants AI to answer booking emails. The team collects 100 real emails and tries a small model and a frontier model. The small one handles routine questions about dates and baggage well, quickly and cheaply. It struggles with complicated complaints. So the agency sends routine emails to the small model and passes complaints to the large one, with a person checking the replies before they go out.
Frequently asked questions
What is a frontier model?
A frontier model is one of the most capable AI models a company offers, at the edge of what AI can currently do. It is usually the slowest and most expensive option in that company's range.
What is a small language model?
A small language model is a lighter AI text model built for focused, less demanding tasks. It costs less to run and answers faster, but it can be less accurate on broad or complex questions.
Should I always use the biggest AI model?
No. Test the models on your own tasks. If a smaller model gets the job right, it will usually be cheaper and faster. Keep the large model for the jobs that really need it.
Can a small model be trained for my business?
Yes. Microsoft says small models can be fine-tuned on a company's own data to do well on a specific task without paying for a larger model.
Related explainers
The DevDrift helps teams pick the right-sized model for each job and test it on real work: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.