What is Granite Embedding Multilingual R2 (311M and 97M)?
Granite Embedding Multilingual R2 is a pair of small open text embedding models from IBM, released on 29 April 2026 under Apache 2.0: granite-embedding-311m-multilingual-r2, with 768-dimension vectors, and granite-embedding-97m-multilingual-r2, with 384. An embedding model turns text into numbers so a search system can match a question to the passages that answer it. They replace IBM's 2024 multilingual R1 models and join the English-only R2 models from 2025.
Both are encoder models on the ModernBERT design with a 262,000-token vocabulary, read up to 32,768 tokens (64 times R1's 512) and cover more than 200 languages, with extra retrieval training for 52 of them and for nine programming languages. The 311M can shorten its vectors to 512, 384, 256 or 128 dimensions. IBM says all training data carries permissive licences and passed its internal data clearance review, and it ships ONNX and OpenVINO builds alongside the PyTorch weights.
Is Granite Embedding Multilingual R2 (311M and 97M) worth it?
Granite Embedding Multilingual R2 earns its place on provenance and long context rather than leaderboard rank. IBM reports 65.2 on multilingual MTEB retrieval for the 311M, 13 points above its R1 predecessor, and 60.3 for the 97M. In IBM's own table of open models under 500 million parameters, the 311M sits second, behind Microsoft's harrier-oss-v1-270m at 66.4 and ahead of Jina v5 text nano at 63.3 and EmbeddingGemma 300M at 62.5.
Long documents are its clearest win. On the LongEmbed benchmark the 311M scores 71.7 against 65.0 for Harrier 270M, 63.6 for Jina's nano and 55.4 for EmbeddingGemma, which makes sense given its 32,768-token window. It is weaker on code and English: Jina's nano model scores 71.2 on MTEB code against Granite's 63.9, and 58.8 on English retrieval against 52.6.
The 97M is the speed option. IBM measures it at 2,534 documents a second on one H100 against 1,828 for the 311M, and it scores 9.4 points above multilingual-e5-small at nearly the same speed. The 311M's shortened vectors hold up well: cutting from 768 to 128 dimensions drops multilingual retrieval only from 63.9 to 62.5 in IBM's table, a sixfold storage saving.
The catch: every comparison here is IBM's, run by IBM's team, and the throughput depends on library versions. IBM notes that Transformers 5.8.0 removed ModernBERT's unpadding optimisation, cutting the 311M from 2,960 to 1,828 documents a second against version 4.57.6. Ollama does not yet run ModernBERT models, and it is text only.
Our take: in regulated companies where training-data provenance is a procurement question, Granite R2 is the easy yes; where only accuracy counts, Harrier's small models are slightly better and also permissively licensed.
Where it earns its place
A knowledge-management lead at a European insurer
Search long policy wordings in a dozen languages without an external API
Embed whole sections up to 32,768 tokens with the 311M on a CPU server via ONNX or OpenVINO; IBM's statement that all training data is permissively licensed shortens the procurement review.
A platform engineer running RAG for 40 internal teams
Keep vector storage costs down across tens of millions of chunks
Store 311M vectors cut to 256 dimensions, a third of the full 768, which in IBM's table costs about 0.4 points of multilingual retrieval accuracy.
A developer adding semantic search to an on-premise document app
Ship an embedder that runs fast on customer hardware without a GPU
Bundle the 97M model with ONNX Runtime; at 384 dimensions and 97 million parameters it is small enough to run beside the application, and Apache 2.0 allows redistribution.
Before you commit
- Setup
- pip install sentence-transformers and load ibm-granite/granite-embedding-311m-multilingual-r2; pass backend='onnx' or 'openvino' to use the bundled optimised builds. FlashAttention 2 is optional.
- Integrations
- Sentence Transformers, Transformers, ONNX Runtime (CPU, CUDA, TensorRT, DirectML), OpenVINO, vLLM embedding endpoint, llama.cpp after GGUF conversion; not Ollama.
- Data and privacy
- Apache 2.0. IBM says training data comes from permissively licensed public pairs, IBM-internal technical data and IBM-generated synthetic data, all through a governance clearance process. Self-hosted, documents stay on your hardware.
- Cost at scale
- Weights are free and the models are small; on one H100 IBM measured 1,828 documents a second for the 311M and 2,534 for the 97M at 512 tokens, so a 10-million-chunk index takes roughly 1.5 hours on the 311M.
- Limits
- Text only; 32,768 tokens per input; vector shortening on the 311M only; speed drops on Transformers 5.8.0 versus 4.57.6; enhanced retrieval training covers 52 of the 200+ languages.
What’s good
- Apache 2.0 with training data IBM says is permissively licensed and cleared
- 32,768-token context; 71.7 on LongEmbed in IBM's tests
- 97M runs at 2,534 documents a second on one H100
- 311M vectors shrink to 128 dimensions with little loss
- ONNX and OpenVINO builds for CPU and Intel hardware
What’s not
- Second to Harrier 270M on multilingual retrieval (65.2 against 66.4)
- Behind Jina v5 nano on code and English retrieval
- All benchmarks are IBM's
- Not supported by Ollama
- Text only
Best for
Enterprises that need a small, multilingual embedder with clear data provenance for search over long documents.
Granite Embedding Multilingual R2 (311M and 97M) alternatives
Granite Embedding Multilingual R2 (311M and 97M): common questions
- Is Granite Embedding Multilingual R2 (311M and 97M) worth it?
- Granite Embedding R2 is the small embedder a legal department will sign off quickly, though Harrier's 270M edges it on accuracy. We rate it 3.9 out of 5.
- How much does Granite Embedding Multilingual R2 (311M and 97M) cost?
- Free under Apache 2.0. The model cards list no hosted endpoint; self-host on CPU or GPU.
- What are the downsides of Granite Embedding Multilingual R2 (311M and 97M)?
- Second to Harrier 270M on multilingual retrieval (65.2 against 66.4); behind Jina v5 nano on code and English retrieval; all benchmarks are IBM's; not supported by Ollama; text only.
- Who is Granite Embedding Multilingual R2 (311M and 97M) best for?
- Enterprises that need a small, multilingual embedder with clear data provenance for search over long documents.
- How much does Granite Embedding Multilingual R2 (311M and 97M) cost for a team or in production?
- Weights are free and the models are small; on one H100 IBM measured 1,828 documents a second for the 311M and 2,534 for the 97M at 512 tokens, so a 10-million-chunk index takes roughly 1.5 hours on the 311M.
- Is Granite Embedding Multilingual R2 (311M and 97M) safe for company data?
- Apache 2.0. IBM says training data comes from permissively licensed public pairs, IBM-internal technical data and IBM-generated synthetic data, all through a governance clearance process. Self-hosted, documents stay on your hardware.
- What are the best Granite Embedding Multilingual R2 (311M and 97M) alternatives?
- The closest alternatives we have reviewed are Harrier-OSS-v1 (270M, 0.6B, 27B), EmbeddingGemma 2, Jina Embeddings v5, Qwen3-Embedding and Qwen3-Reranker and all-MiniLM-L6-v2.
Answers checked Nov 7, 2026.