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

2026-05-04 · Source: Hugging Face - Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Advanced, long

Summary

IBM Granite has released two new Apache 2.0 licensed multilingual embedding models, Granite Embedding Multilingual R2, designed for enterprise use. The models include a 97M-parameter compact version and a 311M-parameter full-size version. Both support over 200 languages, with enhanced retrieval quality for 52 languages and 9 programming languages, and feature an extended 32,768-token context window, a 64x increase over previous R1 models. The 97M model achieves a 60.3 score on MTEB Multilingual Retrieval, outperforming all other open sub-100M multilingual embedders, while the 311M model scores 65.2 and includes Matryoshka embedding support for flexible dimensionality. These models are built on the ModernBERT architecture and are compatible with popular frameworks like `sentence-transformers`, LangChain, LlamaIndex, Haystack, and Milvus.

Key takeaway

For AI Architects and NLP Engineers building multilingual RAG systems or cross-lingual search, the Granite Embedding Multilingual R2 models offer a compelling upgrade. Your teams can significantly improve retrieval quality and context handling for 200+ languages and code, especially for long documents, without complex integration. Consider the 97M model for efficiency or the 311M model for peak performance and Matryoshka flexibility.

Key insights

New IBM Granite multilingual embedding models offer superior retrieval quality and extended context for diverse language and code applications.

Principles

• Compact models can achieve high retrieval quality.
• Longer context windows significantly improve long-document retrieval.
• Matryoshka embeddings enable flexible dimension-quality trade-offs.

Method

Models are trained using knowledge distillation from multiple teachers, contrastive fine-tuning on multilingual retrieval pairs, and model merging, with vocabulary selection for compact versions.

In practice

• Use `granite-embedding-97m-multilingual-r2` for high throughput.
• Truncate 311M embeddings for reduced storage/computation.
• Integrate with LangChain or LlamaIndex via one-line model change.

Topics

• Multilingual Embeddings
• Granite Embedding R2
• ModernBERT Architecture
• Matryoshka Representation Learning
• MTEB Multilingual Retrieval

Code references

• ggerganov/llama.cpp
• ibm-granite/granite-embedding-models

Best for: AI Architect, NLP Engineer, Research Scientist, Machine Learning Engineer, AI Scientist, AI Engineer

Related on AIssential

• End-to-End Evaluation of a RAG System for Hospital Documents in Portuguese — Optimizing RAG components individually enhances performance for querying specialized documents in resource-limited settings.
• FLiP: Towards understanding and interpreting multimodal multilingual sentence embeddings — FLiP models effectively interpret sentence embeddings by linearly extracting lexical content, revealing underlying biases.
• SkMTEB: Slovak Massive Text Embedding Benchmark and Model Adaptation — Specialized text embedding models can be effectively adapted for low-resource languages using existing multilingual foundations.
• When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval — Query embedding interpolation significantly improves multilingual dense retrieval, but English dominance creates an asymmetric benefit.
• Evaluation of Chunking Strategies for Effective Text Embedding in Low-Resource Language on Agricultural Documents — Effective text chunking significantly improves RAG performance in low-resource, morphologically complex languages.
• Google's Gemini Embedding 2 arrives with native multimodal support to cut costs and speed up your enterprise data stack — Gemini Embedding 2 unifies diverse media types into a single semantic space, enhancing AI efficiency and accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hugging Face - Blog.