Model
A learned system that turns inputs into outputs.
Go deeper
A model is not the app, agent, search index, safety policy, or workflow around it.
Technical details
Training establishes parameters; inference applies those parameters to new inputs.
Deep Archive + Fossil Hall
Peel from a plain-language answer into the distinctions that keep model selection accurate over time.
Provider and deployment
The maker of a model and the place it runs are separate facts.
Go deeper
One release can appear through its maker, a cloud platform, or a local runtime.
Technical details
Deployment offers differ in identifiers, regions, controls, latency, limits, and price.
Family and lineage
Related releases share ancestry without being interchangeable.
Go deeper
Family labels overlap; a model can be general, reasoning-oriented, multimodal, and coding-capable.
Technical details
Lineage connects versions while capability tags support nonexclusive classification.
Release, version, and alias
A friendly name may point to a moving target or a fixed snapshot.
Go deeper
Stable aliases ease upgrades; dated releases improve reproducibility.
Technical details
Record canonical model IDs, aliases, release dates, and lifecycle events independently.
Tokens and context budget
Models process bounded pieces of text, images, audio, and output.
Go deeper
A large window is capacity, not proof that stuffing it produces a good answer.
Technical details
Input, retrieved evidence, instructions, tool results, reasoning, and output share finite budgets.
Modality
Text, images, audio, and video are different input and output channels.
Go deeper
Multimodal names do not guarantee every direction or feature is supported.
Technical details
Verify input and output modalities for the exact release and endpoint.
Reasoning
Some models spend more inference work on difficult problems.
Go deeper
More reasoning can improve hard tasks while increasing latency and cost.
Technical details
Reasoning controls and behavior vary by provider and release; evaluate the workload.
Configuration and sampling
Settings shape variability, length, effort, and tool behavior.
Go deeper
A familiar control may be absent or behave differently on a newer model.
Technical details
Treat configuration as release-specific interface data, not universal physics.
Structured output
A schema asks for machine-checkable fields rather than free-form prose.
Go deeper
Valid syntax does not prove that extracted values are correct.
Technical details
Validate the schema, field evidence, retry path, and human review threshold separately.
Embeddings, retrieval, and reranking
Search components can find evidence before a generator writes.
Go deeper
Retrieval is often a better system choice than sending an entire document collection.
Technical details
Embedding, indexing, candidate retrieval, reranking, citation, and generation are distinct stages.
Tools, agents, and systems
A capable model becomes useful through boundaries, tools, state, and checks.
Go deeper
Many use cases need several components rather than one larger model.
Technical details
Model calls, deterministic code, permissions, retrieval, evaluators, and human gates form the system.
Hosted, open-weight, open-source, and source-available
Downloadable weights are not automatically open source.
Go deeper
Licenses, training transparency, source code, and the freedom to run or modify are different dimensions.
Technical details
Record access mode and license claims explicitly; do not infer them from availability alone.
Fossil Hall
No release in the opening census is retired. When a model becomes legacy, deprecated, retired, or unavailable, its dated record will remain here rather than disappear.
PreviewCurrentLegacyDeprecatedRetired