"AI" has come to mean chatbot. Yet most of the models you run into every day write no text at all. Your phone finds "beach" among your photos. Your video call transcribes the meeting. Your bank blocks a suspicious payment. Those are models, but rarely LLMs, large language models.
A task goes in and an output of a different shape comes out: text, image, transcript, search by meaning, a score, then a small model in-house and a voice assistant that chains three models.
A family of models, not just one
Key concept
An LLM is one kind of model among several. Choose the model by the shape of the output you need: a text, an image, speech, a search result, a number or a category. The most impressive model is not always the best fit.
Choose by the output you need
Before you pick a tool, write down in one line what the task must produce. The answer points to a family:
- a text (summarize, draft, extract, answer): an LLM;
- an image or a video: an image or video model;
- text from speech, or speech from text: a speech model;
- finding what looks like a question, a document or a product: embeddings;
- a number or a category (a score, a forecast, a yes or no): classic machine learning, or an LLM constrained to a format fixed in advance;
- any of these outputs, without your data leaving your machine or your company: a small model that runs locally.
Several families often work together in one product. A voice assistant chains a model that transcribes, an LLM that answers and a model that reads the answer aloud.
The model map
For each family: what it does, what it changes for you, a few concrete uses and its main limit.
LLMs: reading and writing
An LLM reads and produces text. Most can also read an image or a scanned document: a screenshot, an invoice, a diagram.
What it changes for you. Instructions are enough: no training data needed. Each answer is paid for in tokens (chunks of text, often parts of words) and takes a few seconds. Your texts go to the provider, unless you host the model yourself.
Uses. Summarize a contract or meeting notes. Draft a first reply to a customer. Extract amounts and dates from invoices in varied formats.
Its limit. It produces the most plausible continuation, not the most accurate one: it can be wrong with confidence (see Hallucinations).
Image and video models: creating visuals
They create or edit an image or a short video from a description, or from a starting image.
What it changes for you. A visual arrives in seconds instead of a photo shoot. As of October 2026, an image costs a few cents at Google, and a few seconds of video less than a dollar. Usage rights and "AI-generated" labels have to be checked with each provider.
Uses. Turn one product photo into several formats and settings. Produce mockups or illustrations to test an idea. Remove an object or extend the background of an image.
Its limit. These are tools for creating, not for deciding. Text inside images, hands, logos and consistency from one image to the next remain fragile: a person reviews before anything is published.
Speech models: transcribing and speaking
They work in both directions. Speech-to-text converts a voice into text. Text-to-speech reads a text in a natural voice. Chained with an LLM, they give you an assistant you can talk to.
What it changes for you. Transcription is billed per minute of audio, and it is cheap. As of October 2026, at OpenAI and Google Cloud alike, an hour-long meeting costs less than a dollar. A synthetic voice makes a text available to people who cannot or would rather not read it. A recording of a person is personal data: tell participants and get their consent.
Uses. Transcribe and summarize a meeting or a customer call. Dictate a report in the field. Read a manual or a message aloud for people with low vision.
Its limit. Accents, background noise, industry jargon and proper names lower accuracy. Test on your own recordings, not on a studio demo.
Embeddings: searching by meaning
Embedding models write nothing. They turn a text, or an image, into a list of numbers, so that two pieces of content close in meaning get close numbers.
What it changes for you. It is very cheap: as of October 2026, a million tokens costs a few cents with OpenAI's smallest model. Search is fast. You compute the numbers once, store them in a database, then compare. This is the search building block of RAG, retrieval-augmented generation (see RAG: look it up before you answer).
Uses. Find the right internal document without knowing the exact word. Suggest similar products or articles. Group thousands of pieces of customer feedback by theme.
Its limit. Close in meaning does not mean correct or up to date. Quality depends on how documents are split. Measure search on your own questions.
Classic machine learning: numbers and categories
This is the oldest family and often the most profitable. A model learns from your past data, usually a table: transactions, customer records, sensor readings. It returns a category, a score or a forecast. Tree-based models still lead in this area.
What it changes for you. It needs labeled examples and someone to train it, then monitor it. After that, it answers in milliseconds, at almost no cost per decision, and usually gives a probability with its answer.
Uses. Score the risk of fraud or late payment. Forecast next week's sales or stock. Send a ticket to the right queue.
Its limit. It only does what it learned. If your customers or products change, its results degrade without warning: it has to be retrained.
For a decision: classifier or LLM?
When the output is a choice among known options, two approaches are worth comparing. The most common is still a classic classifier trained on your past decisions. The other is an LLM with structured outputs. As of October 2026, Anthropic, OpenAI and Google let you constrain the answer to a schema fixed in advance, for example a category taken from a list.
So the format no longer sets them apart. The classifier needs labeled examples to learn from; after that, it answers fast, at a very low cost per decision. The LLM starts without training data and handles a wide range of texts; each answer costs more and takes longer. An effective approach is often to start with an LLM, keep its reviewed decisions as examples, then train a classifier once the volume justifies it. In both cases, the format can be guaranteed; correctness has to be measured.
Very high volume: the classifier costs far less in total, at a very low cost per decision. The volume justifies training a classifier.
Small models, in-house
A small open model, whose weights are published, is downloaded and runs on your own hardware: a laptop, a phone, a company server. Nothing goes to a provider.
What it changes for you.
- Your data stays with you: useful for confidential or medical documents, or covered by professional confidentiality.
- It works offline, in a warehouse, on a construction site or on a plane.
- No bill for each use: you pay for the hardware and the electricity, not for tokens.
- It does not change without you: the model stays the same until you update it.
- In exchange, it is less capable than a large hosted model, especially for multi-step reasoning or long texts. It needs suitable hardware, and you own the updates, the security and the Monday-morning outage.
Uses. Summarize or sort sensitive documents without sending them outside. Classify and tag thousands of messages a day, with no cost per message. Support a field team with no connection. Add a feature to a mobile app that works without a network, such as a suggested reply or a summary. Complete code on a developer's machine. Prototype a feature before choosing a hosted model.
Examples from October 2026. Google publishes Gemma 4 in several sizes, from the smallest for phones to the largest for a personal computer, under the Apache 2.0 license. Mistral publishes its small Ministral 3 models under the same license. Microsoft presents Phi-4-mini as a model to run directly on the device, under the MIT license. For apps, Apple's Foundation Models framework gives access to the model built into the device, which works without a connection. And JetBrains' Full Line code completion runs entirely on the developer's machine, without sending any code over the internet, according to the vendor.
Is a small model enough? Do not guess: measure. Gather about fifty real cases with the expected answer, run them through the small model and a large hosted model, and compare the success rate, the speed and the cost. If the gap is small on your cases, the small model is enough. The Evaluations (evals) page explains how to build that test set.
What it means for you
For founders
Before wiring an LLM into everything, list the repetitive tasks in your product and the output of each. If it is a category or a score, a classic model can cost far less than an LLM and answer faster. If your customers refuse to let their data leave, a small local model can open a market. Your margin depends on it once usage grows.
For product managers
Write down the expected output of each AI feature before you talk about models: a text, an image, a transcript, a ranking, a score. That output sets the cost, the response time and the way you measure quality. A feature that must work offline or in under a second already points to a small model or a classic one.
For business leaders
When a vendor sells an "AI-powered" solution, ask which model runs behind it, and why that one. A precise answer, such as "a classifier trained on your tickets", is a better sign than the name of a famous LLM. For sensitive data, ask where the model runs, and whether a local version exists.
For tech leads
For a closed choice, require a typed output and validate it, whatever the model. Compare candidates on your own data, with latency and cost per decision, not on launch figures. Hosting an open model adds operational load, but keeps data inside your perimeter and freezes its behavior, which often matters more in a large company.
Go deeperWhat calibration means
A model is calibrated when its probabilities match reality: out of a hundred decisions made at 80% confidence, about eighty are right. That is useful for automation. Above a threshold, the decision goes through; below it, a human reviews. A classic classifier usually returns a probability with its decision. An LLM does not always give one, and the one it states is not necessarily calibrated. Whatever the model, check its calibration on your own cases before you set a threshold.
Threshold at 80%: 6 decisions of 12 go through, 6 go to a human for review. A model is calibrated when its probabilities match reality: out of a hundred decisions made at 80% confidence, about eighty are right. Whatever the model, check its calibration on your own cases before you set a threshold.
The classic mistake
fAIling RULE XII
The trap: Put "AI-Powered" on everything
Putting "AI" on a feature does not make it better. If a simple rule or a classic model does the job, an LLM adds cost, latency and unpredictable errors. Pick the tool for the problem, not for the press release.
Read rule XII in the manifestoTry it
List five tasks in your team that could benefit from AI. For each one, write down in one line the expected output, then the family of models that produces it, using the map above. Then take the one that ends in a decision, such as accept, route or prioritize: first test a simple rule on fifty real cases, and count how many it resolves.
Around this topic
Go deeperAn example from October 2026: Jev
TypeSafe AI announced Jev on September 15, 2026, as its first "System One" model: it takes an unstructured situation and returns a decision among options defined in advance, together with a probability. At launch, the vendor claimed a response time of 70 to 500 milliseconds, an input price of $0.042 per million tokens, calibrated probabilities and a model that "can't hallucinate". Those are the vendor's figures, not independent measurements, and the vendor's page still lists Jev as in early access in October 2026.
The last promise deserves a close reading. TypeSafe explains that it guarantees an answer that matches the expected format, not a correct decision: the same guarantee a structured output gives.
Go deeperClassic machine learning still wins
For tabular data, tree-based models remain formidable. A 2022 study across 45 tabular datasets concludes that, on medium-sized data, XGBoost and random forests still beat deep learning. Since then, models pretrained for tables have shaken that finding on small data: TabPFN, published in Nature in January 2025, beats the best methods on datasets of up to 10,000 rows, according to its authors. And the first of Google's published rules of machine learning is not to be afraid to launch a product without machine learning.
Go deeperEmerging families, as of October 2026
These families get a lot of attention, but few readers have a use for them today. Unless you work in the field concerned, watch them before you build on them.
- World models generate an environment that reacts to your actions over time. Google DeepMind presents Genie 3 as a real-time interactive world model. Its prototype, Project Genie, is limited to some subscribers in the United States. Since May 2026, NVIDIA has published Cosmos 3, an open model used to simulate and produce training data. According to NVIDIA, robot makers and a carmaker have already adopted it. For you: a topic for robotics, self-driving cars and video games.
- Vision-language-action models drive a robot from what it sees and an instruction in plain language. Google DeepMind keeps Gemini Robotics 2, announced in July 2026, for partners such as Apptronik and Boston Dynamics. Since April 2026, NVIDIA has published the weights of Isaac GR00T N1.7, built for humanoid robots. For you: worth following if your work involves physical automation.
- Scientific models predict physical or biological phenomena. AlphaFold predicts the shape of proteins. Demis Hassabis and John Jumper, of Google DeepMind, received half of the 2024 Nobel Prize in Chemistry for this work. Its open database, hosted by EMBL-EBI, lists over 200 million predictions. The European Centre for Medium-Range Weather Forecasts (ECMWF) has produced ensemble forecasts with its AIFS model since July 2025. For you: results you will use through services, rather than models to deploy.
- Time-series foundation models forecast a curve (sales, traffic, consumption) without training on your data. Amazon has published Chronos-2 since October 2025, under the Apache 2.0 license. Since August 2026, Google has published TimesFM 3, whose weights are for non-commercial use only, and offers it in BigQuery as a preview. For you: the family closest to a concrete use. Before you change methods, compare these models with your current forecasts on your own history.
What to do with it
Before choosing a model, write in one line what the task must produce: text, image, speech, search result, number or category.
For a category or a score, start with an LLM, keep its reviewed decisions, then train a classifier once volume justifies it.
Before choosing between a small local model and a large hosted one, measure both on about fifty of your real cases.