IBM releases Granite Embedding R2, a multilingual model for semantic search

IBM has introduced Granite Embedding Multilingual R2, an open multilingual model for semantic search under the Apache 2.0 license. On popular benchmarks, it leads the category of models with fewer than 100 million parameters.

What the model is

Embedding models turn text into vectors: sets of numbers that encode the meaning of the text. They underpin RAG (Retrieval Augmented Generation) systems: first, an embedding model finds relevant documents in a database, then a large generative model builds an answer from that information.

Granite R2 supports a context of 32 thousand tokens, four times more than standard embedding models. This means the model can analyze whole chapters and entire documents at once, finding the information needed for search.

Technical specifications

The model was trained on data in more than 30 languages, but it works as a single universal network without special adapters. On MTEB and other standard benchmarks, R2 shows the best performance among all models in its weight class.

The key point is compactness. Fewer than 100 million parameters means the model runs on modest hardware: a developer laptop, a lightweight GPU, or even a CPU is enough:

• 30+ languages in one model
• 32K tokens of context instead of the typical 8K
• Fewer than 100M parameters, so it runs fast on standard hardware
• Apache 2.0 license, which permits commercial use

Why it matters

The choice used to be simple: open lightweight models of mediocre quality, or closed cloud APIs that require an internet connection and money. Granite R2 breaks that pattern.

For enterprises, it means full control over data, since everything runs locally without the cloud. For startups, it means simple integration and cheaper scaling. There is no dependence on provider quotas and no network-request latency.

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

it to specific languages and domains."

What it means

Multilingual embedding models have reached a 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 moving from cloud APIs to local solutions.