Granite Embedding R2
IBM's Apache 2.0 ModernBERT embedding models: 47M to 311M parameters, 8K to 32K context, English or 52 enhanced languages plus code.
In short. Granite Embedding R2 is a set of small embedding models from IBM that turn text and code into numbers for semantic search. They are tiny, from 47 million to 311 million parameters, so they run quickly on a CPU or a small GPU. There are English models and multilingual models, all under Apache 2.0, trained on data that IBM describes as permissive and enterprise-friendly.
Summary
IBM released the second generation (R2) of its Granite Embedding models in two steps: two English models on 15 August 2025 and two multilingual models on 29 April 2026. All four use the ModernBERT architecture and the Apache 2.0 license. The English pair are granite-embedding-english-r2 (149M parameters, 768-dimensional vectors) and granite-embedding-small-english-r2 (47M parameters, 384 dimensions), both with an 8,192-token context. The multilingual pair are a 311M model (768 dimensions) and a 97M model (384 dimensions), with a 32,768-token context.
IBM says the multilingual models cover more than 200 languages through the encoder's pretraining, with explicit retrieval and cross-lingual training for 52 languages and for code (Python, Go, Java, JavaScript, PHP, Ruby, SQL, C and C++). They support Matryoshka dimension reduction, so vectors can be cut to 512, 384, 256 or 128 dimensions, and they ship with ONNX and OpenVINO versions; the card also notes compatibility with vLLM and llama.cpp (GGUF). IBM says the training data uses only permissive, enterprise-friendly licenses plus data that IBM collected or generated, which matters for companies that audit what a model was trained on.
IBM reports that the 311M multilingual model scores 65.2 on the 18-task Multilingual MTEB Retrieval benchmark, 13 points above its predecessor's 52.2. These are IBM's own figures, so compare with other labs' models on the public leaderboard before choosing. The small English model is the most downloaded in the family, with about 6.8 million downloads in the 30 days to 11 November 2026 on Hugging Face. IBM also publishes an English reranker, granite-embedding-reranker-english-r2, to re-score the best results.
Good for
- Fast, low-cost semantic search and retrieval-augmented generation on CPUs and small GPUs
- Multilingual and cross-lingual retrieval, including code retrieval, with the 311M and 97M models
- Long documents of up to 32K tokens (multilingual) or 8K tokens (English)
- Companies that want an Apache 2.0 model with a permissive, documented training-data policy
Watch out for
- Four models with different sizes, languages and context limits, so pick by language and latency
- Quality is best on the 52 enhanced languages and code, while other languages rely on the encoder's pretraining only
- The benchmark scores are IBM-reported on IBM's chosen benchmarks
- Vectors are not interchangeable between models or sizes, so switching means re-indexing
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Granite Embedding R2. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| English small 47M BF16 | 48M | 0.1 GB | 0.115 GB | Laptop or phone |
| Multilingual 97M BF16 | 97M | 0.2 GB | 0.23 GB | Laptop or phone |
| English 149M BF16 | 149M | 0.3 GB | 0.345 GB | Laptop or phone |
| Multilingual 311M BF16 | 312M | 0.62 GB | 0.713 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- BGE-M3 vs Granite Embedding R2: which embedding model should you use?
- Best open-source embedding models in 2026
Frequently asked questions
- What is Granite Embedding R2?
- Granite Embedding R2 is a set of small embedding models from IBM that turn text and code into numbers for semantic search. They are tiny, from 47 million to 311 million parameters, so they run quickly on a CPU or a small GPU. There are English models and multilingual models, all under Apache 2.0, trained on data that IBM describes as permissive and enterprise-friendly.
- Can I use Granite Embedding R2 commercially?
- Yes. Granite Embedding R2 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Granite Embedding R2 open source?
- Granite Embedding R2 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Granite Embedding R2 on my own computer?
- Yes. The smallest variant, English small 47M, needs about 0.115 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Granite Embedding R2 released?
- Granite Embedding R2 was released by IBM on April 29, 2026.
- How long a context window does Granite Embedding R2 have?
- Granite Embedding R2 supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Granite-Embedding-311M-Multilingual-R2 model card (Hugging Face) huggingface.co
- Granite-Embedding-Small-English-R2 model card (Hugging Face) huggingface.co
- Granite Embedding R2 Models (arXiv 2508.21085) arxiv.org
- Granite Embedding repository (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| BGE-M3 BAAI | 568M | 8K | MIT (Permissive) | 2.61 GB Laptop or phone | |
| Qwen3-Embedding Alibaba Qwen | 600M – 7.57B | 32K | Apache 2.0 (Permissive) | 1.38 GB Laptop or phone | |
| EmbeddingGemma 2 Google DeepMind | 270M – 740M | 8K | Apache 2.0 (Permissive) | 0.621 GB Laptop or phone | |
| Jina Embeddings v5 Jina AI | 239M – 677M | 32K | CC BY-NC (Non-commercial) | 0.552 GB Laptop or phone |