TL;DR
IBM has introduced two new multilingual embedding models, granite-embedding-97m-r2 and 311m-r2, supporting over 200 languages and code retrieval. Both models outperform previous versions and are open under Apache 2.0, aimed at enterprise deployment.
IBM has released two new multilingual embedding models under the Apache 2.0 license, designed to support over 200 languages with improved retrieval quality and enterprise readiness. These models, based on ModernBERT architecture, aim to address the trade-off between model size and language coverage in multilingual AI applications.
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
The two models are named granite-embedding-97m-r2 and granite-embedding-311m-r2, with the former being a compact 97 million-parameter model and the latter a full-size 311 million-parameter model. Both support 200+ languages, including 52 with explicit retrieval training, and handle context lengths up to 32,768 tokens, a significant increase over previous versions.
Both models are designed for out-of-the-box use with popular frameworks such as sentence-transformers, transformers, LangChain, LlamaIndex, Haystack, and Milvus, requiring minimal integration effort. They are optimized for CPU inference via ONNX and OpenVINO, making them suitable for deployment in enterprise environments. The models also support cross-lingual code retrieval across nine programming languages, expanding their utility for multilingual coding tasks.
Why It Matters
This release addresses a persistent challenge in multilingual AI: balancing broad language support with model efficiency. By providing high-performance, open-source models under a permissive license, IBM enables organizations to implement multilingual retrieval and search applications without licensing restrictions or extensive infrastructure changes. This could accelerate adoption of multilingual AI solutions across industries, especially in global enterprises.
Products Worth Considering
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Cypress Green
As an affiliate, we earn on qualifying purchases.
Background
Previous models like XLM-RoBERTa offered limited context windows and smaller language support, constraining their effectiveness in long-document and multilingual retrieval tasks. IBM’s earlier R1 models laid groundwork, but the new R2 models represent a significant architectural upgrade, leveraging ModernBERT and enhanced training data to improve performance across benchmarks like MTEB Multilingual Retrieval.
The release follows ongoing industry efforts to improve multilingual NLP, with competitors also releasing models, but IBM emphasizes enterprise readiness, governance, and open licensing as key differentiators.
“These models represent a significant step forward in multilingual retrieval, combining broad language coverage with enterprise-grade performance and open licensing.”
— IBM AI Research Team
“Supporting 200+ languages with high-quality retrieval, especially in long documents and code, opens new possibilities for global organizations.”
— IBM Data Science Lead
What Remains Unclear
It is not yet clear how these models will perform in real-world enterprise deployments at scale, or how they compare with proprietary solutions in specific use cases. Further testing and user feedback are expected to clarify their practical advantages and limitations.
What’s Next
IBM plans to facilitate broader adoption by providing detailed deployment guides and integration support. Future updates may include fine-tuning tools, additional language support, and expanded benchmarking results. Monitoring user feedback and real-world performance will shape subsequent development.
Key Questions
What are the main advantages of these new models?
The models support over 200 languages, handle long context lengths, improve retrieval quality, and are open-source under Apache 2.0, making them suitable for enterprise deployment with minimal integration effort.
How do these models compare to previous versions?
The R2 models outperform their R1 predecessors in retrieval benchmarks, with the 97M model achieving a +12.2 point gain and the 311M model a +13.0 point gain, thanks to architecture improvements and better training data.
Are these models ready for production use?
Yes, both models are designed for enterprise deployment, supporting frameworks like sentence-transformers and transformers, with optimized inference options and broad language support.
What are the licensing terms?
The models are released under the Apache 2.0 license, allowing free use, modification, and distribution, suitable for commercial applications.
Will there be updates or additional languages supported in the future?
IBM plans to expand support and improve performance based on user feedback, with future updates likely including more languages and fine-tuning tools.
