What the parts of a model name actually mean.

Every learned weight a model holds, counted once. It sets what must be stored and loaded, not what runs per token.

The size of the whole checkpoint: all learned weights across every layer and, in a mixture-of-experts model, across every expert. Total parameters govern download size, memory to hold the model, and the hardware needed to load it at a given precision. On a dense model this is also the number that runs on each token; on a sparse model it is not, which is why a name that states only one number can be read two different ways. ModelTree records total and active separately and never collapses them into a single figure.

Also writtentotal paramstotal parameter countparams totalparameter count

RelatedActive parametersMixture of expertsExpert count suffix

The weights actually used to produce each token. On a mixture-of-experts model this is far smaller than the total the checkpoint holds.

The subset of a model's weights that participate in computing any single token. A router selects a few experts per token, so a sparse model computes as though it were much smaller than it is while still needing the whole checkpoint in memory. Active parameters therefore track inference cost and latency, while total parameters track storage and memory. A name like "17B" on a mixture-of-experts release is usually the active figure, which is why reading it as the model's size understates the model by a wide margin.

Also writtenactivated parametersactive paramsactivated paramsparameters per tokenactive parameter count

RelatedTotal parametersMixture of expertsExpert count suffix

An architecture that holds many expert sub-networks and routes each token to only a few, so active parameters fall well below total parameters.

A model built from many parallel expert sub-networks plus a router that picks a small number of them per token. The checkpoint must carry every expert, so the total parameter count stays large, while only the selected experts do work, so the compute per token stays small. This is the single architectural reason a model name can carry two very different parameter numbers, and it is why ModelTree records total and active as separate fields rather than one.

Also writtenmoesparse mixture of expertssparse moeexpert routing

RelatedActive parametersTotal parametersExpert count suffix

The E in a name such as 17B-16E: the number of experts the checkpoint holds. It explains the gap between active and total parameters.

A suffix of the form <number>E in a repository id or model name, counting the expert sub-networks in a mixture-of-experts checkpoint. Paired with the parameter figure it is the clearest signal a name gives that active and total parameters differ: two models can share an active count and differ several-fold in total size purely by expert count. The suffix is a publisher convention rather than a standard, so its absence does not mean a model is dense.

Also written16E128EE suffixexpert suffixxNE

RelatedMixture of expertsTotal parametersActive parameters

The trained weights can be downloaded and run yourself. It says nothing about the licence terms, and it is not the same claim as open source.

A release whose learned parameters are published for download, so the model can be run outside the creator's own service. Availability is the whole of the claim. The licence attached may restrict fields of use, redistribution, or user counts, and the training data and training code may not be published at all. ModelTree stores weight availability and OSI licence approval as two separate booleans for exactly this reason: a model can satisfy the first and fail the second, and most widely used open-weight models do.

Also writtenopen weightsweights availabledownloadable weightsopen-weight model

RelatedOpen source AI

The Open Source Initiative's bar: freedom to use, study, modify and share, plus the data information and code needed to rebuild the parameters.

A defined term rather than a description of mood. Version 1.0 of the Open Source Initiative's Open Source AI Definition grants four freedoms — use, study, modify, share — and requires the preferred form for making modifications, which for a machine-learning system means data information, the complete training and inference code under OSI-approved licences, and the parameters under OSI-approved terms. The definition states explicitly that open source models and open source weights must include the data information and code used to derive those parameters, so publishing a checkpoint is not on its own enough to earn the term.

Also writtenosaidopen source modelopen source ai definitionopen source weights

RelatedOpen weight

How much input and output a model can hold at once, counted in tokens rather than words or characters.

The maximum number of tokens a model can attend to in one request, spanning the prompt, any retrieved material, the conversation so far, and in most accountings the response. Publishers state it in tokens and usually abbreviate — 200K, 1M, 10M — so the unit is easy to lose. Maximum output is a separate, smaller limit stated separately, and a stated window is a capability claim, not a promise of quality at that length. ModelTree stores the token figure rather than the abbreviation, so that a page's 128K and the recorded 128000 cannot drift apart. That figure is the reading the source supports: usually the plain decimal one — 128K as 128000, 1M as 1,000,000 — but the exact integer where a page states one, and a binary size where the page fixes one (Solar Pro Preview's "4K" is 4096, on its Phi-3-4K lineage), so the stored number is never tidier than its source.

Also writtencontext lengthcontext sizemax contextwindow size

RelatedModel API id

The exact string you send to an API. It is a serving-platform identifier, so one model has several — and none of them is the model's name.

The literal identifier a request must carry to reach a model on a given platform. Anthropic's overview lists a distinct id for the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry for the same model; Google's Gemini documentation calls the equivalent field the endpoint. Because the id belongs to the platform rather than to the model, an id is never a safe key for identity across platforms, and ModelTree keeps model, product, and serving platform as separate entities so that one model's several ids do not become several models.

Also writtenapi idmodel idmodel stringendpoint idmodel name string

RelatedModel aliasDated snapshotContext window

An id that points at whichever release is current. Convenient, and it means the model behind your code can change without your code changing.

An identifier that resolves to a moving target rather than to one fixed snapshot. Google documents a latest alias that "will get hot-swapped with every new release of a specific model variation", with two weeks' notice by email before a breaking change; Anthropic lists an alias row separately from the API id row for each model. An alias is the right choice when you want the newest behaviour and the wrong one when you need a result to be reproducible, because the same request can reach different weights on different days.

Also writtenaliaslatest taglatest aliasfloating aliasrolling pointer

RelatedModel API idDated snapshotPreview version

A date inside a model id, pinning one fixed release. It is what makes a request reproducible when an alias would not be.

A model identifier carrying an explicit release date, so it resolves to one set of weights and keeps resolving to them. Publishers write the date in several shapes — a trailing YYYYMMDD, an @-separated date, an embedded MM-YYYY — and none of these is standard across vendors. A dated snapshot is the counterpart to an alias: it is what you pin when a result must be reproducible, and its date is a release marker, not a training-data cutoff or a verification date.

Also writtensnapshotsnapshot idpinned versiondate suffixdated version

RelatedModel aliasModel API idPreview version

A release track marked in the id itself. Google documents stable, preview, latest and experimental as four distinct tracks with different guarantees.

A marker inside a model id naming which release track the model is on rather than what the model is. Google's documentation sets out four: stable, which usually does not change and is recommended for production; preview, which may be used in production but carries tighter rate limits and at least two weeks' deprecation notice; latest, an alias that is hot-swapped; and experimental, which is not suitable for production and whose endpoints may disappear. The track is a lifecycle fact, so ModelTree records it as lifecycle status on the release rather than reading it out of the name.

Also writtenpreviewpreview modelexperimental versionstable version

RelatedModel aliasDated snapshot

The -Instruct on a repository id marks the instruction-tuned build. The same release without it is the pretrained base model, and they behave differently.

A suffix distinguishing the instruction-tuned checkpoint from the pretrained base one within a single release. Meta ships both and states the difference plainly: instruction-tuned models are meant for assistant-like chat and visual reasoning, while pretrained models are for adaptation to natural language generation. The two share a family, a parameter count, and almost all of a name, so the suffix is often the only thing in an id that tells you which artefact you have — and prompting a base model as though it were tuned is a common consequence of missing it.

Also writteninstructinstruction tunedinstruction-tuned modelchat tuned

RelatedModel API idExpert count suffix

A tag such as Q4_K_M or F16 naming the numeric precision weights are stored at. It changes file size and output, never the parameter count.

A marker on a converted checkpoint recording how its weights were re-encoded at lower precision. Hugging Face's GGUF documentation lists the types and their arithmetic: F16 is a 16-bit IEEE 754 half-precision float, while Q6_K is 6-bit quantization in super-blocks of sixteen blocks of sixteen weights, resolving to 6.5625 bits per weight. Because quantization is applied after training by whoever converted the file, a quantized build is a distribution artefact rather than a new model, and its tag is not a claim the original creator made.

Also writtenquantizationquantisationQ4_K_Mquant tagbits per weight

RelatedTotal parameters