Private beta

Unless you pick otherwise, a tool hands you a small, quick, low-cost model, because most questions do not need more. It starts writing immediately. For rewording an email, shortening a paragraph, or summarizing a document you pasted in, the speed is worth having and the quality is fine.

Models labeled reasoning or thinking write out a chain of intermediate steps before they show you anything. Working step by step measurably improves accuracy on problems that take several steps (Wei et al., 2022), which is what an eligibility rule with three conditions is, or a comparison between two documents. They are slower, sometimes by a minute, and they cost more to run.

The steps a model shows you are not a record of how it reached the answer. Researchers planted a hint in the prompt that changed what the model answered, and the model then wrote a confident step-by-step explanation that never mentioned the hint (Turpin et al., 2023). A reasoning model makes fewer careless errors. It can still be confidently wrong, and you still verify anything you will act on.

Use the reasoning model when a mistake would cost your client something, when the question holds more than one condition, or when you want the tool to check work instead of produce it. The fast model is enough for tone, for drafts, and for text you were going to read closely anyway.

Training adjusts billions of internal numbers until the model's guess at the next token matches the text it is learning from. Those numbers are the parameters, and they are all the model keeps. Counting them, 8 billion or 400 billion, measures how much pattern the model has room to hold. It is not a count of facts, because the model stores none.

Three inputs decide how capable a model turns out. How much text it learned from, how much computation went into training it, and how many parameters it has. Quality improves smoothly with all three, and raising one alone wastes the other two (Kaplan et al., 2020). Train a large model on too little text and a smaller model trained on more text will beat it (Hoffmann et al., 2022). Companies that sell closed models publish none of the three, so even the claim that one model is bigger than another usually cannot be checked from outside.

A large language model, hundreds of billions of parameters, runs in a data center and reaches you over the internet. A small language model, a few billion parameters, runs on a laptop or a phone with no internet connection at all. Small models are cheap, fast, and private, and they are noticeably more limited. Both are built on purpose now, for different jobs.

A small model will not be good at everything. Trained further on one narrow job it can hold its own against a much larger generalist on that job, and a large model can be used to teach a small one, a method called distillation (Hinton et al., 2015). It will not generalize past its lane. Inside that lane it often does the work at a fraction of the computation and the electricity, which matters for cost and for the footprint.

Tokenizers cut English into few pieces and many other languages into far more, up to five times as many for the same content, so the same request costs more, runs slower, and fills the model's working memory faster in your clients' languages (Ahia et al., 2023). Shrink a model and the languages it saw least are the first to suffer. A small model that reads well in English is often visibly worse in Vietnamese, Amharic, or Haitian Creole.

The best known assistants are closed. The company keeps the trained parameters, and you reach the model through its app or its programming interface. You cannot inspect it, cannot run it yourself, and cannot keep it when the company changes or retires it. The company also does not have to tell you what the model was trained on, how large it is, or what changed in the last update, and mostly it does not. Your text goes to its servers to be answered.

Other companies publish the trained parameters, so anyone can download the model and run it on their own computer or their agency's server. This is why a laptop with no internet connection can still answer you, and why a small vendor can build a product without renting a model from anyone.

Published weights are the finished numbers. The training data and the training code that produced them usually stay private, so nobody outside can study the model or rebuild it, and the Open Source Initiative reserves the term open source for models that release enough to do that (Open Source Initiative, 2024). Licenses differ too. Some open-weight models carry standard permissive licenses, and others carry custom terms restricting who may use them and for what. Read the license before your agency builds on one.

A model running on hardware your agency controls is the only arrangement in which client text never leaves the building, and that is a real reason to care about open weights. It is also easy to overstate. An app can be built on an open model and still send every word to a cloud server. Ask where the text goes, and get the answer in writing.

These models learned from an enormous amount of web text, along with licensed and purchased material. Most closed models never publish the list, so you cannot check whether the model saw anything good on the rule you are asking about. Web text is heavily English and thin in most of your clients' languages, and the language gap starts there.

Training stops on a date and the model's trained memory stops with it. Many tools now run a live web search to reach past that date, which helps and moves the question. The answer is then only as good as the pages the tool happened to pull, so ask which sources it used and open them.

A fact placed at the start or the end of a long input is found far more reliably than the same fact buried in the middle, and accuracy falls as the input gets longer (Liu et al., 2024). Paste a forty-page policy, ask one narrow question, and the answer can miss a paragraph the model technically read. Quote the section you care about instead of handing over the whole file.

Once a model can predict text, people tune it with human feedback so that it follows instructions, answers in a certain tone, and refuses certain requests (Ouyang et al., 2022). That stage, and not size, is why two models feel different in conversation. It also rewards agreement. Assistants trained this way tend to match the view a user states, and they will abandon a correct answer when the user pushes back (Sharma et al., 2024). Hand the model a wrong premise and expect it to build on it.

Models are retrained and updated, sometimes quietly, and the same name in the menu can point to different machinery than it did last month. Re-test a saved prompt after an update, along with the judgment you formed about what this tool gets right.