Models

Every model in this snapshot of the ibm-granite catalogue (see scope), grouped by family. Within a group, rows are sorted by parameter count. A dash means the field is not published in the model's configuration.

Contents

1. Language — dense — 51
2. Language — sparse (mixture of experts) — 9
3. Language — hybrid (state-space + attention) — 12
4. Language — sliding-window attention — 2
5. Language — adapter switching — 3
6. Code — 15
7. Vision–language — 7
8. Vision retrieval — 1
9. Speech — 9
10. Text embedding and reranking — 12
11. Adapters — 23
12. Time series — 9
13. Speculative decoding heads — 6
14. Earth observation — 7
15. Dataset and classifier releases — 6

Language — dense — 51

Text models in which every layer is the same. The nine GneissWeb.7B_ablation_* rows are listed here by architecture only: they are data-ablation releases — the same 7B model trained on different datasets, three seeds each — and not general-purpose Granite models. They belong with the dataset and classifier releases.

_{↑ contents}

Language — sparse (mixture of experts) — 9

A router sends each token to a few of many expert MLPs.

_{↑ contents}

Language — hybrid (state-space + attention) — 12

State-space layers carry most positions, attention layers a minority.

_{↑ contents}

Language — sliding-window attention — 2

Most attention layers see a fixed span rather than the whole context.

_{↑ contents}

Language — adapter switching — 3

A base model that swaps low-rank adapters mid-sequence.

_{↑ contents}

Code — 15

Trained for programming languages; two distinct architecture generations.

_{↑ contents}

Vision–language — 7

An image tower feeding a language model.

_{↑ contents}

Vision retrieval — 1

Document-image embeddings for retrieval.

_{↑ contents}

Speech — 9

Speech recognition and speech-conditioned language models.

_{↑ contents}

Text embedding and reranking — 12

Encoders for retrieval, and cross-encoders that rerank.

_{↑ contents}

Adapters — 23

Low-rank adapters and adapter libraries over a published base.

_{↑ contents}

Time series — 9

Forecasting and reconstruction; a separate lineage from the language models. Their window is measured in past samples, not tokens, so the token-context column is blank for all of them: the published windows are 512 samples for the mixers and PatchTST, 2048 for FlowState and 8192 for PatchTST-FM.

_{↑ contents}

Speculative decoding heads — 6

Small heads that propose tokens for a larger base model.

_{↑ contents}

Earth observation — 7

Geospatial foundation models.

_{↑ contents}

Dataset and classifier releases — 6

Data-ablation artifacts rather than general models: the GneissWeb classifiers and filters, none of which publishes a configuration. The nine GneissWeb 7B ablation language models from the same release publish one and are listed under Language — dense by architecture; together these are the 15 dataset-ablation releases that support.md marks out of scope.

_{↑ contents}