What Is an AI Model?

Headlines announce a new model. You open an app. Those are two different things, and a lot of confusing sentences about AI come from treating them as one.

An AI model is the trained part. An app is the product built around it. One model can sit inside several products at once, and one product can change which model it uses without telling you much about it.

Lathic covers what a large language model is and what an algorithm is separately. This article is about the boundary.

What a model actually is

Ordinary software is written. Somebody typed the rules, and you could in principle read them.

A model is not written that way. It is produced: a learning process runs over a large body of data, adjusting a very large set of numbers until the result performs the task well. Those numbers are the model, usually called parameters, or weights.

You do not have to take a vendor's word for this. Under the EU AI Act, whose obligations for general-purpose model providers apply from 2 August 2025, such a model is exempted from two of its provider's documentation duties, unless it carries systemic risk, only when it is released under a free and open-source license and "parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available." That is the legislature listing the parts: weights, architecture, usage notes. No source code, no data.

The documentation rules agree. A general-purpose model's technical file must state "the architecture and number of parameters," and separately describe "the data used for training, testing and validation" and "the computational resources used to train the model." A model is treated throughout as the output of a production process, not as a program.

The scale is published, and it is not always the number people quote. Meta documents two models of one family whose totals are 109 billion and 400 billion parameters, while noting that "at any point in time, the number of parameters actually doing the compute ('active parameters') on a given token is always 17B." Two very different numbers describe each of them.

Parameter counts are a poor quality score in any case. The AI Act uses one billion only as an indicative signal, and only alongside a second condition: a model with at least a billion parameters "and trained with a large amount of data using self-supervision at scale" should, it says, "be considered to display significant generality." It introduces the test with "could, inter alia." A threshold for breadth, not a ranking.

The law draws the line for you

This is the passage the whole article rests on, and it is worth quoting in full. Recital 97 of the AI Act:

"Although AI models are essential components of AI systems, they do not constitute AI systems on their own. AI models require the addition of further components, such as for example a user interface, to become AI systems. AI models are typically integrated into and form part of AI systems."

A model on its own is not something a person can use. No window, no button, no memory of you, no way to reach the internet. Someone builds all of that around it.

The Act separates the two deliberately, "to enable legal certainty." An AI system is "a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments." A general-purpose AI model is "an AI model" that "displays significant generality" and "can be integrated into a variety of downstream systems or applications."

One caution. The AI Act never defines the bare term "AI model." It defines the system, and it defines the general-purpose model, and uses "AI model" throughout without a definition. Anyone who tells you the EU defines an AI model is quoting something that is not there.

Other bodies also define the system rather than the model, in non-identical words. ISO/IEC 22989 calls an AI system an "engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives." NIST describes "an engineered or machine-based system," and says it adapted that from the OECD and the same ISO standard. Three authorities, three phrasings, one shape.

What the product does that the model does not

The clearest way to see the difference is to follow one request through a product whose maker has published the steps. Microsoft's architecture documentation for its workplace assistant gives the sequence: a user enters a prompt; the product "preprocesses the input prompt by using grounding" and reaches into the organization's own data; it then "sends the grounded prompt to the LLM"; the model generates a response; and the product "returns the response to the app and the user."

Count what happened outside the model. The interface. The retrieval of your documents. The permission check: the documentation states the product "only accesses data that an individual user is authorized to access." The assembly of a prompt you never saw. The delivery of the answer back into your app.

The model did one step of that, the step Lathic covers in what AI inference is.

Safety layers sit outside it too. GitHub's documentation states that for its default models, "input prompts and output completions run through GitHub Copilot's content filters for harmful, offensive, or off-topic content." That filter belongs to the product, which is why one model can behave differently in two apps.

It is also why "the AI searched the web" and "the AI remembered our last conversation" are loose sentences. The product did those things and handed the results to the model as text.

Three things people merge, kept apart

The algorithm and the model are separate defined terms in the standards literature, and EU law refers to a system's capability "to derive models or algorithms, or both, from inputs or data." Two things, not one.

Training data is the one people merge most often, and the law keeps it apart. A general-purpose model's provider must "draw up and make publicly available a sufficiently detailed summary about the content used for training," separately from anything about the model itself, and the technical file must describe the data's "type and provenance" and "curation methodologies." You do not describe a thing that is the same thing.

Be careful of the overcorrection. It is not true that nothing of the data survives. NIST's adversarial machine learning taxonomy catalogs "training data extraction" as an attack category precisely because some of it can sometimes be recovered. A model is a compressed and lossy product of its data, and pulling the data back out is an attack rather than a feature.

One more boundary. The Act's recitals say the definition of an AI system should not cover software based on "the rules defined solely by natural persons to automatically execute operations." Nothing was learned, so there is no model.

Why several products can use the same model

Because that is the arrangement the industry is built on, and the law was written expecting it. A general-purpose model provider's technical file must record "the tasks that the model is intended to perform and the type and nature of AI systems in which it can be integrated," and providers must hand documentation to "providers of AI systems who intend to integrate the general-purpose AI model into their AI systems." The Act describes a supply chain because a supply chain is what exists.

Three documented examples.

Amazon's hosting service provides "access to high-performing foundation models from leading AI companies." In its code samples the model is chosen by a string in the request, and the same call with a different string reaches a different company's model. Microsoft's catalog advertises models "from Microsoft, Azure OpenAI, Anthropic, DeepSeek, Meta, Mistral, Cohere, Hugging Face, and more," and its July 2026 update states that it "includes over 10,000 models," running "from foundation models, reasoning models, small language models, multimodal models, domain-specific models, and industry models."

One model provider can be reached through several clouds at once. Google Cloud offers another company's models on its own platform as "fully managed and serverless models as APIs," and Anthropic notes that its published retirement dates apply to its own platforms, and that "partner-operated platforms (Amazon Bedrock and Google Cloud) set their own retirement schedules, so a model's lifecycle status and dates can differ." The same model, five front doors, not all on the same retirement calendar.

And the model inside a consumer chat product is often purchasable separately. OpenAI publishes an identifier described as pointing to "the latest Instant model currently used in ChatGPT."

Why one product can change its model

The reverse is just as common. GitHub's assistant is the clearest published case: its documentation says the product "supports multiple AI models, each with different strengths," and its reference table lists models from several competing providers. It warns users that "model availability is subject to change. Some models may be replaced or updated over time," and publishes its own retirement history, on its own dates, separate from those set by the companies that made the models.

It also documents choosing for you. Its automatic selection "routes the task to the optimal model," reserving "higher-cost reasoning models for problems that truly need it, while routing straightforward tasks to faster, lower-cost models." That is a model router, and the engineering detail is revealing: routing happens "along natural cache boundaries" to avoid extra cache costs, and "switching models mid-session has shown increased cost without ample improvements in quality."

So when a product feels different one week to the next, there are three ordinary explanations before any exotic one. It switched models. Its prompt or retrieval changed. Or the name stayed the same while the thing behind it moved.

What a model name actually promises

A model name can mean two opposite things, and only the documentation tells you which.

It can be a fixed snapshot. Anthropic states that each of its model identifiers "identifies a pinned version of the model," that "the underlying model remains constant for the lifetime of that ID," and that it "maps to a single, fixed model snapshot." It names the common error directly: the belief that a dateless identifier behaves "as evergreen pointers that route to the latest or best-performing version" is, in its words, "not the case." The same page notes that this covers model IDs and not the convenience aliases its API still accepts for some earlier models, which resolve to the most recent dated snapshot for that minor version. One vendor, both behaviors, documented.

Whichever it is, the version is usually documented in a model card. The idea comes from a 2019 peer-reviewed paper proposing that released models carry documentation of how they perform and what they are for, and it is now routine: one large hosting platform calls model cards "files that accompany the models and provide handy information."

Or it can be a pointer that moves. Google publishes a version convention of stable, preview, latest and experimental, and says of the latest alias that it "will get hot-swapped with every new release of a specific model variation," with two weeks' email notice before a breaking change.

Same-looking convention, opposite behavior. A model version string is not self-explanatory, and a reader who assumes one vendor's rule applies to another has no way of knowing which they have got.

One smaller caveat. Even a pinned model is served by infrastructure that changes: Anthropic notes that "occasionally, infrastructure updates produce minor differences in observable behavior even when the model ID and weights have not changed." Not everything that shifts is a new model.

When a model goes away

Models retire on published schedules, and the vocabulary is precise enough to matter.

Anthropic publishes four states. Active means "fully supported and recommended for use." Legacy means the model "will no longer receive updates and may be deprecated in the future." Deprecated means "still functional but no longer recommended," with a named replacement and a retirement date. Retired means "no longer available for use," and "requests to retired models will fail."

Deprecation is therefore a warning, not a switch-off. Retirement is the switch-off, and it is hard: requests past that date fail rather than degrade.

Notice periods are published. Anthropic commits to "at least 60 days' notice before model retirement for publicly released models." OpenAI publishes tiered minimums, which it can shorten where "safety or compliance concerns require a faster timeline": "at least 6 months" for generally available models, "at least 3 months" for specialized variants, and as little as "2 weeks" for preview models. It also fixes the vocabulary, using "sunset" and "shut down" interchangeably, and "legacy" for models that "no longer receive updates."

One vendor is unusually candid about the cost. Anthropic writes that it retires models "to ensure capacity for new model releases," and lists what that costs: users who valued a model must migrate, "researchers lose access to models for ongoing and comparative studies," and retirement "introduces safety- and model welfare-related risks." It has "committed to long-term preservation of model weights."

Note what that last commitment implies. The weights can be preserved while the service is switched off, because the weights are the model and the service is not.

What to take away

The model is the trained artifact. The product is everything built around it so a person can use it. The AI Act's recitals state it outright: models "do not constitute AI systems on their own."

Once that is fixed, a lot of confusing sentences resolve. A new model release is not a new app. A product changing its behavior may not be a new model. Two apps giving different answers may be running the same model behind different prompts, retrieval and filters. And a model name may be a promise that nothing will change, or that it will, depending on whose name it is.