A frontier model is one of the most capable AI models available right now, built by a large lab and rented to you over the internet. An open-source model, more precisely an open-weight model, is one whose trained insides are published so anyone can download it, run it on their own hardware, and look at how it's built. The trade is top capability you rent against a slightly older capability you can hold, and for most small businesses in 2026 that gap is a matter of months.
What is a frontier model?
"Frontier" is the industry's word for the leading edge, the handful of models that score best on hard tests at any given moment. They come from labs like OpenAI, Anthropic, and Google, and you never hold one. You type into ChatGPT, Claude, or Gemini, your request travels to the lab's servers, the model answers, and the model itself stays put. That's why people also call these closed models, and why "open source vs. closed source AI" is the same argument in different words.
The label is a moving target. Whatever leads this fall gets passed by spring, so version numbers here would date fast. A CPA firm that picks a tool because it runs "the newest model" will find the newest model has changed before tax season ends.
Open-weight vs. open source, in plain terms
Weights are the numbers a model learned during training. They are the model, in the same way a recipe card is the dish once you have a kitchen. Publish the weights and anyone with a capable enough computer can run the model, adjust it for their own work, and keep their data on their own machine. Meta's Llama family, Alibaba's Qwen models, DeepSeek's models, and OpenAI's gpt-oss release are the names that come up most.
Open source is a stricter claim, and the people who coined the term have written it down. The Open Source Initiative's Open Source AI Definition asks for three things together, the weights, the code used to train and run the model, and enough information about the training data that a skilled person could build something substantially equivalent. Almost no popular model publishes all of that. The training data is the part labs keep, for competitive and legal reasons, so what you can usually download is the finished model without the full recipe for making it.
The license matters too, and it varies. OpenAI put gpt-oss under Apache 2.0, a permissive license that lets you use and modify it commercially. Meta's Llama license is its own document, with an acceptable-use policy and a clause that sends the very largest companies to Meta for separate terms. A salon owner will never hit that ceiling, but it's a reminder to read the license before you build on a model, the same way you'd read a lease.
Frontier and open-weight models compared
Side by side, the two deals differ on four practical points.
| Frontier | Open-weight | |
|---|---|---|
| How you get it | Rent access through an app or a paid connection | Download the weights and run them yourself |
| What you pay | A monthly subscription or a per-use charge | Nothing for the model, plus the hardware, setup, and upkeep |
| Can you inspect it | No, the model stays on the lab's servers | Yes, the weights are yours to examine and adjust |
| Where it runs | The lab's data centers | Your computer, your server, or a cloud machine you choose |
The "what you pay" row hides the real cost. A downloaded model arrives with no chat window, no memory between conversations, and no connection to your files. Those come from the program around the model, and with an open model somebody has to supply it, either you or a person you pay. The weights are free in the way a donated engine is free.
How far behind are open models?
Less than people assume, and the distance moves. Epoch AI, a research group that tracks model capability over time, found in May 2026 that the best open-weight models have trailed the top closed models by about four months on average since January, a distance of about eight points on its Epoch Capabilities Index. Its earlier measurement put the lag nearer three months and noted that the gap has at times nearly closed before opening again, usually when a lab releases something new. Nobody can tell you yet whether four months is where it settles, and the answer changes how long renting stays the obvious default.
Four months is not much for the work a small business actually hands to AI. A model that led the field last winter can draft a guest-reply template for a lodge, summarize a stack of vendor invoices for a retailer, or turn a crew's job notes into a clean estimate for a roofing company. Where the newest frontier models pull ahead is the hardest reasoning, long multi-step tasks, and work across images and documents at once. If your job lives there, the gap shows, and for most office work it won't.
Which should a small business use?
Start with where your data is allowed to go, and let capability settle the rest.
If nothing you'd type is sensitive, rent. A retail shop writing seasonal sale emails gets the best model available, a finished app, and no hardware to babysit for the price of a subscription. That's the right answer for most businesses I'd expect to read this, and it's where my own default sits.
If some of what you'd type can't leave the building, the open-weight side starts to make sense, because it's the only arrangement where the model runs entirely on hardware you control. Patient records, client files under a confidentiality duty, and anything a contract says stays in-house belong in that group. Whether you actually need it, or whether a business plan with a signed data agreement covers you, is the subject of do you need private AI. What a local model is and how it answers without the internet is in what a local LLM is.
There's a third answer that more businesses end up with, which is both. A rented model handles the everyday drafting, and an open one handles the files that can't go out. Sending each job to the model that fits it has a name, and model routing explains how it works. Where any of this sits in the bigger picture of how far to take AI in a business is laid out in how to use AI in your business.