Granite Embedding English R2
ibm-granite/granite-embedding-english-r2
published Jul 2025 · updated Jan 2026
Granite Embedding English R2 is a embed model that generates high-quality text embeddings for retrieval, search, and similarity applications.
specs
| Task | Embedding |
| Architecture | ModernBERT bi-encoder |
| Parameters | 149M |
| License | Apache 2.0 |
about this model
Granite-embedding-english-r2 is a 149M parameter dense bi-encoder embedding model that generates fixed-length vector representations (768 dimensions) for text inputs, enabling comparison via cosine similarity for retrieval and search applications. It supports a context length of up to 8,192 tokens and is built on the ModernBERT architecture, incorporating alternating attention lengths, rotary position embeddings, GeGLU activations, and Flash Attention 2.0 for efficiency.
The model is trained exclusively on open-source relevance-pair datasets with permissive, enterprise-friendly licenses, plus IBM-collected and generated synthetic data. It does not use the MS-MARCO dataset due to its non-commercial license. Training data was filtered to remove hate, abuse, and profanity.
Key Strengths
- Strong performance across diverse retrieval domains: text (BEIR, MTEB-v2), code (CoIR), long-document search (MLDR, LongEmbed), conversational multi-turn (MTRAG), and table retrieval (NQTables, OTT-QA, AIT-QA, MultiHierTT, OpenWikiTables).
- Measurable speed advantages of 19–44% over leading competitors while maintaining superior accuracy.
- Training incorporates code in 9 languages (Python, Go, Java, JS, PHP, Ruby, SQL, C, C++).
Benchmark Results
The following table compares the r2 model against its predecessor and other open-source models on key benchmarks:
| Model | Parameters | Embedding Size | BEIR Retrieval (15) | MTEB-v2 (41) | CoIR (10) | MLDR (En) | MTRAG (4) | Encoding Speed (docs/sec) |
|---|---|---|---|---|---|---|---|---|
| granite-embedding-125m-english | 125M | 768 | 52.3 | 62.1 | 50.3 | 35.0 | 49.4 | 149 |
| granite-embedding-english-r2 | 149M | 768 | 53.1 | 62.8 | 55.3 | 40.7 | 56.7 | 144 |
| e5-base-v2 | 109M | 768 | — | — | 50.3 | 32.5 | 37.0 | 115 |
| bge-base-en-v1.5 | 109M | 768 | — | — | 46.6 | 33.5 | 38.8 | 116 |
| gte-modernbert-base | 149M | 768 | — | — | 71.5 | 46.2 | 36.8 | 142 |
Improvements over the previous r1 model include: BEIR +0.8, CoIR +5.0, MLDR +5.7, and MTRAG +7.3.
The model is released under Apache 2.0 license. For further details, see the paper and repository.
best for
- ·Enterprise document retrieval and search
- ·Code retrieval across multiple programming languages
- ·Long-document and conversational multi-turn retrieval
FAQ
It is best for dense retrieval tasks such as enterprise search, code retrieval, long-document search, table retrieval, and multi-turn conversational retrieval.
Use the OpenAI-compatible endpoint on gigarouter with your API key, specifying the model name granite-embedding-english-r2 and sending your text inputs.
It is released under the Apache 2.0 license, allowing unrestricted research and commercial use.
It supports up to 8192 tokens and produces 768-dimensional embeddings.
It is a bi-encoder that generates separate embeddings for queries and passages, compared via cosine similarity.
We're benchmarking and onboarding Granite Embedding English R2 as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.