IBM Granite Multilingual Embedding Sets Retrieval Record

TL;DR: IBM’s Granite Embedding Multilingual R2 family provides the highest-performing open-source multilingual embeddings under 100 M parameters, with the 97 M model scoring 60.3 on MTEB. Both models support a 32 K token context window and cover over 200 languages, making them ready for Thai and other Southeast Asian language search tasks under an unrestricted Apache 2.0 license.

Key facts

• Granite Embedding Multilingual R2 offers two open-source bi-encoder models under Apache 2.0: a 311 M-parameter model and a 97 M-parameter compact version.
• The 311 M model generates 768-dimensional vectors, supports Matryoshka representation learning, and scores 65.2 on the MTEB Multilingual Retrieval benchmark, ranking second among open models under 500 M parameters.
• The 97 M model outputs 384-dimensional vectors and achieves a 60.3 MTEB score, the highest reported for any open multilingual embedder under 100 M parameters.
• Both models cover more than 200 languages, improve retrieval quality for 52 languages, and support code search across nine programming languages.
• They accept inputs up to 32,768 tokens, a 64-fold increase over the previous R1 release’s 512-token window.
• Training uses a five-stage pipeline (retrieval-oriented pretraining, contrastive fine-tuning, knowledge distillation, model merging) built on ModernBERT, with permissively licensed public corpora and IBM-generated data vetted for privacy.
• The models achieve strong results on benchmarks such as BEIR, ClapNQ, COIR, MLDR, LongEmbed, MTRAG, and several table-retrieval suites, demonstrating enterprise-grade performance.

Overview

IBM has added two new models to its Granite Embedding family, targeting enterprise-grade dense retrieval across many languages. Both models are released under the permissive Apache 2.0 license, meaning developers can use them in commercial products without additional legal hurdles. The key selling points are a dramatic increase in context length—up to 32,768 tokens—and the best reported retrieval quality for any open multilingual embedder under 100 M parameters.

Model Variants and Performance

The family consists of a 311 M-parameter model and a 97 M-parameter compact model. The larger model generates 768-dimensional vectors and supports Matryoshka representation learning (a technique that lets the model output vectors of different sizes) [4]. It reaches a score of 65.2 on the MTEB Multilingual Retrieval benchmark, placing it second among open models with fewer than 500 M parameters [4].

The compact model produces 384-dimensional vectors. It was created through model pruning and vocabulary selection, allowing it to retain high retrieval quality while shrinking to 97 M parameters. Its MTEB score of 60.3 is the highest ever reported for an open multilingual embedder under 100 M parameters [4].

Both models cover more than 200 languages, improve retrieval for 52 of those languages, and support code search across nine programming languages [4]. This breadth makes them especially relevant for Southeast Asian markets where language diversity is high.

Why 32K Context Matters

A major architectural change is the 32,768-token context window, a 64-fold jump from the previous R1 release’s 512-token limit [2][4]. In practical terms, the model can embed an entire article, a long product manual, or a multi-turn chat conversation in one go, eliminating the need to split text into short chunks, which often leads to loss of context and poorer ranking. For enterprises that store large documents or need to retrieve information from lengthy support tickets, the larger window translates directly into more accurate search results.

Implications for Thai and Southeast Asian Language Retrieval

Thai and other regional languages have historically lagged behind English in open-source retrieval tools. Models like mBERT or multilingual BGE provide broad language coverage but usually score lower on Thai-specific benchmarks. Granite’s coverage of over 200 languages, combined with its top-ranked scores for 52 languages—including Thai—means developers can achieve better semantic search without resorting to costly proprietary APIs.

The 32 K token capacity also benefits Thai, where sentences can be longer due to the script’s lack of spaces. Being able to process whole paragraphs at once preserves meaning that might otherwise be fragmented.

Licensing and Deployment

Both models are Apache 2.0-licensed, allowing unrestricted commercial use, modification, and redistribution. This is a stark contrast to many other high-performing embeddings that are released under more restrictive licenses. The models are available on Hugging Face, and they integrate seamlessly with the sentence-transformers and Transformers libraries, making deployment on-premises or in the cloud straightforward.

Training Pipeline and Data

The models are built on a bi-encoder architecture (a model that encodes queries and documents separately) that uses the ModernBERT backbone. Training follows a five-stage pipeline: retrieval-oriented pretraining, contrastive fine-tuning, knowledge distillation, and model merging [1][3]. IBM used permissively licensed public corpora together with its own curated datasets, all vetted for privacy and profanity concerns.

The community has noted that IBM’s blog mentions the larger model was trained on roughly 2 trillion tokens from sources such as Wikipedia and IBM-generated code data, underscoring the scale of the effort [6].

Benchmark Results

The models perform strongly on a suite of retrieval benchmarks, including BEIR, ClapNQ, COIR, MLDR, LongEmbed, MTRAG, and several table-retrieval datasets. These results indicate the models are ready for real-world enterprise use cases, from document search to retrieval-augmented generation and code search.

Conclusion

IBM’s Granite Embedding Multilingual R2 family fills a critical gap for developers in Thailand and the broader Southeast Asian region. By offering the best open-source retrieval quality under 100 M parameters, supporting a 32 K token window, and providing an unrestricted Apache 2.0 license, the models enable high-performance multilingual semantic search without the cost or legal complexity of proprietary solutions. Developers can download the models from Hugging Face today and integrate them into existing pipelines using familiar Python libraries.

Sources

4. Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality (huggingface.co) — 2026-05-14
5. ibm-granite/granite-embedding-311m-multilingual-r2 · Hugging Face (huggingface.co) — 2022-05-26
6. GitHub - ibm-granite/granite-embedding-models (github.com) — 2024-12-10
7. Granite Embedding Multilingual R2 Models (arxiv.org)
8. Granite Embedding R2: Setting New Standards for Enterprise Retrieval (www.linkedin.com) — 2025-10-14