🤗Hugging Face Blog•Stalecollected in 5m
IBM Releases Granite Embedding Multilingual R2 with 32K Context
💡Best-in-class retrieval for sub-100M models with 32K context support—perfect for efficient, large-scale RAG.
⚡ 30-Second TL;DR
What Changed
Open-source Apache 2.0 license for commercial flexibility
Why It Matters
This release provides developers with a highly efficient, high-performance embedding model that reduces infrastructure costs while maintaining accuracy. Its long context support makes it ideal for RAG applications involving large technical manuals or legal documents.
What To Do Next
Integrate Granite Embedding Multilingual R2 into your RAG pipeline to test if its 32K context window improves retrieval accuracy for your long-form datasets.
Who should care:Developers & AI Engineers
Key Points
- •Open-source Apache 2.0 license for commercial flexibility
- •Industry-leading retrieval performance for models under 100M parameters
- •Extended 32K context window for processing long-form documents
- •Optimized for multilingual retrieval tasks
🧠 Deep Insight
Web-grounded analysis with 16 cited sources.
🔑 Enhanced Key Takeaways
- •The release includes two multilingual embedding models:
granite-embedding-97m-multilingual-r2(97M parameters, 384-dimensional embeddings) and a largergranite-embedding-311m-multilingual-r2(311M parameters, 768-dimensional embeddings with Matryoshka support). - •The 97M parameter model achieved a Multilingual MTEB Retrieval score of 59.6 across 18 tasks, outperforming the next-best open multilingual embedding model under 100M parameters (multilingual-e5-small) by 8.7 points.
- •The models support over 200 languages, with enhanced retrieval capabilities for 52 specific languages and various programming codes including Python, Go, Java, JavaScript, PHP, Ruby, SQL, C, and C++.
- •The R2 models are built on the ModernBERT architecture, incorporating advanced features like alternating attention, rotary position embeddings, Flash Attention 2.0 for efficiency, and a compact, purpose-trained 180K-token tokenizer.
- •IBM ensures enterprise readiness by training the models on permissively licensed, commercial-friendly datasets, including IBM-collected and IBM-generated data, and notably excludes the non-commercially licensed MS-MARCO dataset.
📊 Competitor Analysis▸ Show
| Model Name | Parameters | Context Window (tokens) | Multilingual Support | License/Availability | Key Benchmarks (MTEB Retrieval) | Embedding Dimensions | Pricing (per 1M tokens) |
|---|---|---|---|---|---|---|---|
| IBM Granite Embedding Multilingual R2 (97M) | 97M | 32,768 | 200+ (52 enhanced) | Apache 2.0 (Open-source) | 59.6 (Multilingual MTEB) | 384 | Free (self-host) |
| IBM Granite Embedding Multilingual R2 (311M) | 311M | 32,768 | 200+ (52 enhanced) | Apache 2.0 (Open-source) | 65.2 (MTEB-v2 Retrieval Avg) | 768 (Matryoshka: 768, 512, 384, 256, 128) | Free (self-host) |
| Qwen3-Embedding-0.6B | 0.6B | 32,000 | 100+ | Apache 2.0 (Open-source) | 0.8168 (nDCG@3) | Up to 1024 (flexible) | Free (self-host) |
| Qwen3-Embedding-8B | 8B | 32,000 | 100+ | Apache 2.0 (Open-source) | 70.58 (MTEB Multilingual, mid-2025) | Up to 4096 (flexible) | Free (self-host) |
| EmbeddingGemma-300M | 300M | 2,048 | Multilingual | Proprietary (Google DeepMind) | 0.8706 (nDCG@3) | Up to 3072 (adjustable) | $0.15 (Vertex AI) |
| Voyage AI voyage-3-large | N/A | 32,000 | Multilingual | Proprietary (API-only) | Outperforms OpenAI/Cohere on 100 datasets | 1024 | $0.18 |
| Cohere embed-v4 | N/A | N/A | 100+ | Proprietary (API-only) | Leads on multilingual retrieval benchmarks | N/A | $0.10 |
| BGE-M3 (BAAI) | N/A | 8,192 | 100+ | Apache 2.0 (Open-source) | Strongest for multilingual hybrid search | N/A | Free (self-host) |
| Jina Embeddings v3 | N/A | 8,192 | 62+ | CC-BY-NC-4.0 (Open-source) | N/A | 1024 (flex to 32) | Free (self-host) / API from $0.02 |
| OpenAI text-embedding-3-small | N/A | 8,192 | N/A | Proprietary (API-only) | Near top of MTEB for English | 1536 (Matryoshka) | $0.02 |
🛠️ Technical Deep Dive
- Architecture: The Granite Embedding Multilingual R2 models utilize a bi-encoder architecture based on the ModernBERT framework.
- Core Enhancements: Key architectural improvements from ModernBERT include alternating attention lengths for accelerated processing, rotary position embeddings (RoPE) for extended sequence length, and streamlined parameters by eliminating unnecessary bias terms.
- Efficiency: The models integrate Flash Attention 2.0 to improve memory usage and throughput during inference.
- Tokenizer: A compact, purpose-trained multilingual tokenizer with a 180K-token vocabulary is used, derived from the GPT-OSS vocabulary, to maintain broad language coverage at a reduced size.
- Model Variants: The release includes two multilingual models: the 97M parameter model (384-dimensional embeddings) and the 311M parameter model (768-dimensional embeddings).
- Model Pruning: The 97M model is a reduced-size variant, derived from the 311M model through layer pruning (22 to 12 layers) and vocabulary selection, followed by continued distillation training to recover quality.
- Activation Functions: The 97M model uses SiLU activation, while the full-size 311M model uses GeGLU.
- Matryoshka Representation Learning (MRL): The 311M multilingual model supports MRL, allowing embeddings to be truncated from 768 dimensions down to 512, 384, 256, or 128 while preserving quality.
- Training Methodology: Training involves retrieval-oriented pretraining, knowledge distillation from multiple teacher models, and contrastive fine-tuning to ensure strong alignment between query and passage embeddings.
- Deployment: The models are released with ONNX and OpenVINO support and are compatible with vLLM for flexible deployment.
🔮 Future ImplicationsAI analysis grounded in cited sources
IBM's focus on smaller, high-performing open-source models will increase enterprise adoption of on-premise or edge AI solutions.
The 97M model's efficiency and strong performance for its size, combined with the Apache 2.0 license, reduces compute costs and deployment friction for businesses.
The 32K context window will significantly enhance the capabilities of RAG systems for complex enterprise documents.
Processing long-form legal, financial, or technical documents without extensive chunking will lead to more accurate and comprehensive retrieval-augmented generation.
IBM will continue to expand its Granite family with more specialized embedding models or multimodal capabilities.
The existing Granite family already includes language, vision, speech, and guardian models, indicating a strategy of comprehensive AI offerings.
⏳ Timeline
2023-09-07
IBM Granite series of decoder-only AI foundation models announced.
2024-05-06
IBM open-sources four variations of Granite Code Models under Apache 2.0.
2024-10-21
IBM releases Granite 3.0, including dense and Mixture of Experts (MoE) LLMs, designed for function calling.
2025-08-15
IBM introduces English Granite Embedding R2 models with an 8192-token context length.
2026-04-29
IBM releases the Granite 4.1 family, including the Multilingual R2 embedding models, scaling retrieval to 200+ languages and 32K context.
📎 Sources (16)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Hugging Face Blog ↗