Granite Embedding

weights

open(four ungated Apache-2.0 R2 embedding checkpoints, 47M to 311M)

data

closed(permissively licensed public datasets plus select proprietary IBM data, neither enumerated nor shipped)

code

partial(usage and documentation in ibm-granite/granite-embedding-models

license

Apache-2.0(OSI)

Apache-2.0 on every R2 checkpoint, ungated. IBM leans hard on data provenance in its own prose - permissively licensed public sources, enterprise-friendly - but the claim stops short of the ladder's test: the mixture is characterized rather than enumerated, it includes proprietary IBM data, and no dataset or script ships. The product repository is documentation and usage, so code is partial and the family lands at the open-weights rung.

• https://huggingface.co/ibm-granite/granite-embedding-english-r2/raw/main/README.md recorded 2026-09-11

  license: apache-2.0 in the front matter and named again under Model Details; training data described as permissively licensed open-source plus select proprietary data, with no dataset listing; the repository is given as ibm-granite/granite-embedding-models.

• https://raw.githubusercontent.com/ibm-granite/granite-embedding-models/main/README.md recorded 2026-09-11

  The repository introduces the collection, links the Hugging Face models and IBM documentation, and states the Apache-2.0 offer; it carries no pretraining or contrastive-training pipeline.

• https://huggingface.co/api/models/ibm-granite/granite-embedding-english-r2 recorded 2026-09-11

  gated: false; private: false; cardData license apache-2.0.

6,764,831 downloads in the trailing 30 days across the granite-embedding R2 embedding checkpoints. granite-embedding-reranker-english-r2 is excluded - every other vendor in this category has its reranker scored as a separate product, and counting it here would apply a different rule to IBM. It adds about 24K and moves no band. The superseded R1 line is also excluded.

• https://huggingface.co/api/models/ibm-granite/granite-embedding-small-english-r2 recorded 2026-09-11

  downloads: 6324429; gated: false; private: false.

• https://huggingface.co/api/models/ibm-granite/granite-embedding-311m-multilingual-r2 recorded 2026-09-11

  downloads: 117691; gated: false; private: false.

• https://huggingface.co/api/models/ibm-granite/granite-embedding-97m-multilingual-r2 recorded 2026-09-11

  downloads: 99229; gated: false; private: false.

• https://huggingface.co/api/models/ibm-granite/granite-embedding-english-r2 recorded 2026-09-11

  downloads: 52733; gated: false; private: false.

Rung 3, the solid workhorse. The numbers are credible and published per checkpoint, and R2 genuinely leads its size class on the enterprise-shaped tasks IBM built the line for - table retrieval, long documents, conversational multi-turn. But the scope is narrower than the frontier by design: the strongest checkpoint is 149M and English-only at 62.8 MTEB-v2, against the anchor's 74.3 multilingual, and the multilingual R2 models are smaller still. Placed against the rung definition rather than a peer: no single measurement reports this product and a category peer together - MTEB v2 English over 41 tasks against the anchor's MMTEB v2 multilingual mean - and rule (d) forbids an edge without one.

• https://huggingface.co/ibm-granite/granite-embedding-english-r2/raw/main/README.md recorded 2026-09-11

  Evaluation table reports granite-embedding-english-r2 at 53.1 BEIR retrieval (15), 62.8 MTEB-v2 (41), 55.3 CoIR and 56.7 MTRAG, with the small-english-r2 variant at 50.9 / 61.1, and a second table showing it ahead of gte-modernbert-base and snowflake-arctic-embed-m-v2.0 on the enterprise-shaped average.