NVIDIA launches Nemotron ColEmbed V2 for AI search
NVIDIA introduces Nemotron ColEmbed V2, a family of multimodal models that searches text, tables, charts, and images. Its 8B version leads the ViDoRe V3 benchmark, but it requires storing many more representations than single-vector models.
NVIDIA has introduced Nemotron ColEmbed V2, a family of models that searches for information in documents where text is only part of the content. It can also find answers in tables, charts, images, scanned pages, and infographics.
The family comes in three sizes: 3B, 4B, and 8B. Its goal is to improve multimodal retrieval, meaning the ability to locate relevant content even when the question is written as text and the answer appears inside an image or a visually complex page.
The model leading the benchmarks
The nemotron-colembed-vl-8b-v2 ranks first on ViDoRe V3, a benchmark designed to measure how systems retrieve useful pages from business documents. It scored 63.42 on NDCG@10, a metric that evaluates whether the most relevant results appear among the first ten.
The other models in the family also rank near the top of their respective size categories:
nemotron-colembed-vl-4b-v2: third place, with a score of 61.54.llama-nemotron-colembed-vl-3b-v2: sixth place, with a score of 59.79.
According to NVIDIA, all three models achieve state-of-the-art results on the V1, V2, and V3 versions of ViDoRe. The comparison reflects results available on February 3, 2026.
What changes compared with traditional search
Many systems convert an entire query and each document into a single vector, a numerical representation that makes it possible to compare meanings quickly. It is efficient and takes up little storage, but it can lose important details.
Nemotron ColEmbed V2 uses another approach known as late interaction. Instead of storing a single representation for each document, it generates multiple vectors for its different elements. It then compares each part of the query with every part of the document and combines the strongest matches.
This makes it possible to distinguish, for example, between a page that mentions a figure and another that contains the exact table answering the question. It also helps connect a text query with a chart or image, even when there is no literal word match.
The trade-off is clear: you have to store many more vectors for the entire document corpus. That is why these models are designed for cases where accuracy matters more than minimizing resource use.
Designed for more visual RAG systems
The technology can be used in RAG systems, applications that first retrieve relevant information and then pass it to a generative model to produce an answer. In practice, you could ask about a company’s revenue and receive a report page containing a chart, a table, or both as the source.
It can also be used for:
- Multimedia search engines.
- Search across documents and other visual formats.
- Assistants that understand full pages, charts, and infographics.
- Business systems that query scanned reports or documents with complex layouts.
NVIDIA distinguishes this family from its previously introduced Llama-Nemotron-Embed-VL-1B model. The 1B model prioritizes low storage use and a high volume of queries, while ColEmbed V2 is aimed at maximum accuracy and keeps multiple representations of each document.
What is inside the models
The models use bidirectional attention, which allows them to analyze the entire input sequence instead of processing it in only one direction. They also apply a ColBERT-based mechanism that calculates similarity between elements of the query and elements of the document.
The 3B model is built from a combination of SigLIP 2 and Llama 3.2. The 4B and 8B versions are based on Qwen3-VL-4B-Instruct and Qwen3-VL-8B-Instruct, respectively. NVIDIA also added multilingual synthetic data, hard-negative examples that look relevant but are not, and a model-merging technique to improve stability without adding latency during inference.
For you, the difference will not appear as a new interface, but in the quality of answers that depend on complicated documents. The next point to watch is whether the improvement in testing justifies the higher storage and processing costs when these models move from experiments into real business systems.