IBM released Granite Embedding R2 — a multilingual model for semantic search

IBM presented Granite Embedding Multilingual R2 — an open-source multilingual model for semantic search under the Apache 2.0 license. According to tests on popular benchmarks, the development leads in the category of models with fewer than 100 million parameters.

What is this model

Embedding models transform text into vectors — sets of numbers that store information about the meaning of the text. This is the foundation for RAG systems (Retrieval Augmented Generation): first, such a model finds relevant documents in a database, then a large generative model creates an answer based on this information. Granite R2 supports a context of 32 thousand tokens — four times more than standard embedding models. This means that the model can simultaneously analyze entire chapters and whole documents, finding the information needed for search.

Technical characteristics

The model was trained on data in more than 30 languages, but works as one universal network without special adapters. Based on test results on MTEB and other standard benchmarks, R2 shows the best performance among all models in its weight class. The main advantage is compactness. Fewer than 100 million parameters means the model works on simple hardware: a developer's laptop, a lightweight GPU, or even a CPU is enough:

• 30+ languages in one model
• 32K token context instead of typical 8K
• Less than 100M parameters — fast on standard hardware
• Apache 2.0 license — commercial use permitted

Why is this needed

Previously, the choice was simple: open lightweight models with mediocre quality or closed cloud APIs that require internet and money. Granite R2 breaks this stereotype. For corporations, it means complete control over data — everything works locally without the cloud. For startups — simple integration and cheaper scaling. No dependence on provider quotas, no delays from network requests.

"Open development means that the community can improve the model and

adapt it to specific languages and domains."

What does this mean

Multilingual embedding models have reached that level of maturity where they are convenient to use in real projects. For developers of RAG systems, this means fewer dependencies on cloud giants and more flexibility in integration. The industry is gradually transitioning from cloud APIs to local solutions.