IBM Granite Embedding English R2

Granite-Embedding-English-R2 turns a piece of English text into 768 numbers. Documents that mean similar things land near each other in that space, which is how site search, document retrieval and retrieval-augmented generation find the right paragraph before a language model ever sees it. The model does not talk, does not answer questions and produces no text at all. It is the larger of the two English models IBM released in August 2025, and on the company's own measurements it wins nearly everything: 53.1 on BEIR retrieval against 50.9 for the 47-million-parameter version, 55.3 against 53.8 on code retrieval, 40.7 against 39.8 on long documents, and 56.7 against 48.1 on multi-turn conversational retrieval — the widest gap in the table, worth almost nine points. It pays for that with speed: 144 documents per second on one H100 against 199 for the small model. And almost nobody downloads it. In the thirty days to 1 September 2026 the repository was pulled about 59,000 times, against 5.75 million for granite-embedding-small-english-r2. The better model is chosen roughly once for every ninety-eight times the faster one is. That ratio is the clearest statement of what production retrieval actually optimises for: an embedding model runs on every indexed document and every incoming query, so throughput and memory decide the choice long before a benchmark table does. The architecture is ModernBERT — 22 layers, 12 attention heads, an intermediate size of 1,152, GeGLU activations, rotary positions, alternating attention windows, a 50,368-token vocabulary trained jointly on code and prose, and no bias terms. Context is 8,192 tokens. It replaces granite-embedding-125m-english, the previous generation, which it beats while running about as fast. IBM's comparison table is unusually candid about where it loses. Against gte-modernbert-base, a rival of identical size, Granite trails badly on code retrieval — 54.8 against 71.5 — and wins the overall average only because it is far stronger on conversational retrieval, 57.6 against 36.8. Readers picking a model for a code search product should read that row before the average. Training data excluded MS MARCO, the retrieval corpus most open embedding models are built on, because its licence forbids commercial use; the encoder was pretrained on IBM's GneissWeb. The licence is plain Apache 2.0 with no territorial carve-out and no acceptable-use annex. English only — Polish text needs the multilingual R2 model instead.

▸Videos

No videos yet.

▸Reviews