There is a legal problem quietly sitting inside many AI search systems, and most businesses have no idea it is there.
A dataset called MS-MARCO, which Microsoft built from real Bing search queries, has been used to train a large number of the most popular free embedding models available today. The problem is that MS-MARCO is explicitly restricted to non-commercial use only. IBM noticed this and built their new Granite R2 models without touching it. Several competing models, including widely-used defaults in popular AI frameworks, carry this risk. If your legal team ever audits your AI stack, this is the kind of detail that surfaces.
IBM's new release covers two models. The smaller one has 97 million parameters and is now the top-ranked free multilingual search model in its size class. The larger one has 311 million parameters and sits second among all free models under 500 million parameters. Both are released under Apache 2.0, the cleanest and most widely understood commercial license in software. Enterprise legal teams already know it, so no special review is needed.
The practical improvement over the previous IBM models is large. The previous version could only read about 512 words at a time before cutting the document off. The new models handle up to 32,000 words in one pass. For a legal contract, a technical manual, or a multi-page research report, this is the difference between the AI reading the whole document and the AI judging a 50-page contract by its cover page. On the benchmark that specifically tests long documents, the smaller model improved by 31 points over its predecessor.
The models also handle code written in nine programming languages including Python, Java, and SQL. This matters for any company building internal tools where developers or analysts are searching across both documentation and actual code.
What makes these models particularly interesting for non-technical buyers is how they fit into existing tools. They work as a direct swap into LangChain, LlamaIndex, Haystack, and Milvus, which are the most common frameworks teams use to build AI search applications. A one-line change replaces an old model with this one. If a vendor or internal team is currently using an English-only default, that same one-line change gives the system coverage across 200 languages.
The commercial embedding market tells you who these models are competing with. Cohere's enterprise multilingual API is considered the standard for production multilingual search at scale, but you cannot run it on your own servers, the pricing is opaque, and your documents travel through their infrastructure. OpenAI's embedding API is the most widely deployed but was not designed for multilingual use as a primary feature. Both of those are paid API services with no self-hosting option.
The IBM models run inside your own infrastructure. Your documents never leave your systems. For industries where data residency or regulatory compliance matters, which includes most of insurance, healthcare, finance, and legal, this distinction alone is enough to make them the obvious choice for evaluation.
The larger 311-million-parameter model also supports a feature where you can compress the search data it produces, reducing storage requirements by 3x or 6x while losing almost nothing in search quality. This matters when you are indexing millions of documents. A 6x reduction in storage is a meaningful cost saving at scale.
The open-source embedding space is moving fast, and IBM is not alone here. Microsoft's Harrier model and Jina AI's offerings are also serious competitors. But the combination of clean licensing, no data leaving your building, strong multilingual performance, and compatibility with tools teams are already using puts IBM's new models near the top of what any serious evaluation should include.