Glossary
AI model release glossary
Plain-language definitions for the terms that show up when AI models are released, updated, or changed — and why each one matters for understanding a change.
Glossary terms
- AI model
A trained system that takes an input and produces an output — such as text, code, or images — based on patterns learned from data.
Why it mattersThe model is the unit of change Modelmark tracks. A release, update, or deprecation is almost always about a specific model or its variants.
Related: Model family · Frontier model · Open weights
- AI model release
The point at which a new model, or a new version of a model, is made available — often via an announcement, model card, or API listing.
Why it mattersA release is the most common kind of model change. Recording when and where it was announced anchors everything else to a source.
Related: AI model update · Provenance · Source-backed record
- AI model update
A change to an existing model — improved behaviour, a new checkpoint, revised limits, or documented capability changes — without a wholly new model name.
Why it mattersUpdates are easy to miss because the name may stay the same. A record makes a quiet behaviour change visible and checkable.
Related: AI model release · Capability signal · Benchmark claim
- Model family
A group of related models that share a lineage or design, usually offered in several sizes or tiers under one name.
Why it mattersChanges often apply to one member of a family but not others. Naming the family and the variant keeps a record precise.
Related: AI model · Frontier model
- Frontier model
A model at or near the leading edge of current capability, typically the largest or most capable in its family.
Why it mattersFrontier releases draw the most claims and the most hype, which is exactly where source discipline and confidence labels matter most.
Related: Model family · Benchmark claim · Reasoning model
- Open weights
A model whose trained parameters are published for download, so it can be run, inspected, or fine-tuned outside the provider's API.
Why it mattersOpen-weight releases change how and where a model can be used. The license and availability terms are part of the change worth recording.
Related: AI model release · Fine-tuning · Inference
- Context window
The maximum amount of text — measured in tokens — a model can consider at once, including both the input and its response.
Why it mattersContext-window changes are a frequent, practical update. They affect what a model can be used for, so they belong in the record.
Related: Token · AI model update
- Token
A chunk of text — often a word piece — that a model reads and generates. Pricing and context limits are usually measured in tokens.
Why it mattersBecause pricing and limits are stated per token, token terms are how many model changes are actually quantified.
Related: Context window · Inference
- Inference
Running a trained model to produce an output from an input — the act of using a model, as opposed to training it.
Why it mattersAccess, pricing, and rate-limit changes are about inference. Distinguishing inference from training keeps a record accurate.
Related: Token · Open weights · MCP
- Fine-tuning
Further training of an existing model on additional data to adapt its behaviour for a specific task or domain.
Why it mattersWhen a provider adds, changes, or removes fine-tuning support, it changes what builders can do — a recordable access or capability change.
Related: Open weights · AI model update
- Structured outputs
A model feature that constrains responses to a defined format, such as valid JSON matching a schema.
Why it mattersStructured-output support is a capability change builders depend on. Its arrival or change is worth recording with a source.
Related: Tool use · Capability signal
- Tool use
A model's ability to call external functions, APIs, or tools as part of producing an answer.
Why it mattersTool-use changes affect what agents can be built. They are capability signals that should be separated from marketing claims.
Related: Structured outputs · MCP · Reasoning model
- Reasoning model
A model designed to spend additional computation working through a problem step by step before answering.
Why it mattersReasoning variants are often released alongside standard ones. Recording which is which avoids conflating different models.
Related: Frontier model · Tool use · Benchmark claim
- Model deprecation
A notice that a model is being phased out — still available for now, but scheduled to be removed or no longer recommended.
Why it mattersDeprecations have deadlines. Capturing the notice and its source gives teams time to plan a migration.
Related: Model retirement · AI model update
- Model retirement
The point at which a model is fully removed from service and can no longer be called.
Why it mattersRetirement is a hard change that can break products. A dated, sourced record is the difference between planning and surprise.
Related: Model deprecation · Inference
- Benchmark claim
A stated result on a standardized test used to describe a model's performance, such as a score on a reasoning or coding benchmark.
Why it mattersBenchmark claims are interpretation, not raw fact. Modelmark keeps them separate from what verifiably changed.
Related: Capability signal · Confidence label · Fact vs interpretation
- MCP (Model Context Protocol)
An open protocol that lets AI applications and agents connect to external tools and data sources through a standard interface.
Why it mattersMCP is how agents read context. Modelmark's MCP surface is meant to return model-change records with their provenance attached.
Related: Tool use · Provenance · Source-backed record
- Provenance
The origin and history of a piece of information — where it came from and how it can be traced back to its source.
Why it mattersProvenance is the core of a trustworthy record. Every Modelmark change is meant to travel with the provenance behind it.
Related: Source-backed record · Confidence label · MCP
- Confidence label
A marker on a record indicating how well-supported it is — for example official, corroborated, single-source, or inferred.
Why it mattersNot all information is equally certain. Confidence labels let you weigh a change honestly instead of treating every claim the same.
Related: Provenance · Benchmark claim · Source-backed record
- Source-backed record
A stored entry about a change that points back to the original material supporting it, with timestamps and a confidence label.
Why it mattersIt is what Modelmark produces. A record you can check and return to is more durable than a feed you have to keep up with.
Related: Provenance · Confidence label · AI model release
Keep reading
See how these terms come together when a change becomes a source-backed record.