Which Granite Embedding Multilingual R2 model should I use?

Use the granite-embedding-97m-multilingual-r2 model for applications where computational efficiency and a smaller memory footprint are critical. It provides the best retrieval performance for its size class. Use the larger granite-embedding-311m-multilingual-r2 model when you need the highest possible retrieval quality and can accommodate the larger model size. Its Matryoshka support also offers the flexibility to truncate embedding dimensions to balance performance and cost.

Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality

IBM Releases Compact Multilingual Embedding Model with 32K Context

IBM has released two new open-source Granite Embedding Multilingual R2 models under an Apache 2.0 license, directly addressing a persistent challenge in enterprise AI: achieving broad language support without requiring large, computationally expensive models. The release's main highlight is the granite-embedding-97m-multilingual-r2, a 97-million-parameter model that scores 60.3 on the MTEB Multilingual Retrieval benchmark, a figure that surpasses all other open models in the sub-100M parameter class. Both new models feature a 32,768-token context window, a 64x increase over their predecessors, enabling more effective processing of long-form documents for applications like retrieval-augmented generation (RAG) and cross-lingual search.

Under the Hood: ModernBERT and Enterprise-Focused Training

The R2 models represent a significant architectural update, moving from the previous generation's XLM-RoBERTa base to ModernBERT. This shift provides native support for rotary position embeddings, enabling the long context window without interpolation, and incorporates optimizations like Flash Attention 2.0. IBM also notes its training data was selected with enterprise use in mind, using curated public and internal datasets while intentionally avoiding sources with non-commercial licensing like MS-MARCO. This focus on data governance is designed to reduce downstream risk for commercial deployments.

granite-embedding-97m-multilingual-r2: 97M parameters, 384-dimensional embeddings, and a 60.3 MTEB Multilingual Retrieval score.
granite-embedding-311m-multilingual-r2: 311M parameters, 768-dimensional embeddings with Matryoshka support, and a 65.2 MTEB score.
Shared Features: 32K context length, Apache 2.0 license, support for over 200 languages (tuned for 52), and code retrieval across 9 programming languages.

Implications for the Open-Source Ecosystem

This release provides a direct, high-performance upgrade path for developers using popular open-source frameworks. The models are designed as drop-in replacements for LangChain, LlamaIndex, and others, allowing for a one-line model name change to instantly add robust multilingual capabilities. The performance leap on long-document benchmarks is particularly notable, with the 97M model gaining over 31 points on the LongEmbed benchmark compared to its R1 version. This improvement makes the model far more practical for real-world multilingual workloads involving legal contracts, technical manuals, and research papers, which were previously truncated by smaller context windows.

IBM's release of the Granite R2 embedding models, particularly the 97M variant, is a direct play for the enterprise developer market. By prioritizing a permissive Apache 2.0 license, avoiding legally ambiguous training data, and delivering a massive performance uplift in a compact size, they are offering a compelling, low-friction upgrade path for production RAG systems that have been constrained by smaller, less capable, or more restrictively licensed multilingual models.

>> Verify Original Transmission at Hugging Face