What it is
Nemotron ColEmbed V2 is a family of visual document retrieval models from NVIDIA, released in January 2026 in 3B, 4B and 8B sizes. You index each document page as an image, without OCR or layout parsing, and search it with a text question. The 8B model is built on Qwen3-VL-8B-Instruct and has about 8.8 billion parameters; the 3B is built on Llama 3.2.
It is a late-interaction model in the ColBERT and ColPali style: instead of one vector per page, it outputs a 4,096-dimensional vector for every token, and scores a match by comparing each query token with its best-matching page token. Images are split into up to eight 512-by-512 tiles plus a thumbnail, at 256 tokens per tile.
None of the sizes can be used commercially. The 4B and 8B are licensed CC BY-NC 4.0; the 3B uses NVIDIA's own non-commercial licence, which limits use to research and education.
Our review
Nemotron ColEmbed V2 is a research model with production-grade accuracy and a licence that keeps it out of production. NVIDIA's card reports an average NDCG@10 of 63.54 on ViDoRe v3 for the 8B, first on that benchmark when it was published on 26 January 2026, plus 92.65 on ViDoRe v1 and 65.16 on v2, where it sat third behind Tomoro's ColQwen3 8B. The 4B scores 61.42 on v3 and the 3B 59.70. These are NVIDIA's figures, produced with the public MTEB harness, so they can be reproduced, but the ranking has moved since January.
Storage is the bill that arrives later. The 8B emits a 4,096-dimensional vector for every token, and a page can use nine tiles of 256 tokens. By our arithmetic from the card's figures, that is up to about 19 MB a page in BF16, or roughly 19 TB for a million pages before any pooling or quantisation. Tomoro's Apache-licensed ColQwen3 8B uses 320 dimensions and at most 1,280 tokens a page, which its card puts at about 0.82 TB per million images. The scoring step also needs a vector store with multi-vector (MaxSim) support, which rules out many single-vector setups.
The licence closes the commercial door at every size. The 4B and 8B are CC BY-NC 4.0, and the 3B, built on Llama 3.2, sits under NVIDIA's own non-commercial licence, which restricts use to 'non-commercial research and educational purposes' while reserving commercial use for NVIDIA itself. The card lists no NIM container or hosted endpoint, and no Hugging Face inference provider serves it.
Running it is a narrow path too: Linux only, tested on A100 and H100 cards, flash-attn pinned to version 2.6.3 and remote code enabled. The 8B weights are about 17.5 GB. For a team asking whether OCR plus a text embedder is good enough for an archive of scans, it is a useful ceiling to measure against, and the per-task spread on the card, from 79.30 on computer-science pages to 50.84 on physics, says the answer will depend heavily on the documents.
Where it earns its place
An information-retrieval researcher at a university lab
Benchmark a new visual retriever against a strong open baseline
Load the 8B in Transformers and run mteb 2.7 or later on ViDoRe v3; NVIDIA's per-task scores give a reproducible target. Academic use is exactly what the CC BY-NC licence allows.
An ML engineer at an insurer with millions of scanned claim forms
Decide whether the OCR pipeline is worth keeping
Index a few thousand sample pages with the 8B as an accuracy ceiling and compare it with OCR plus a text embedder. The licence rules out deploying it, so the result informs a choice between OCR and an Apache-licensed ColQwen model.
A data-platform lead sizing a multi-vector index
Estimate storage before committing to late interaction
At up to 2,304 tokens of 4,096 BF16 values a page, a 100,000-page archive needs up to about 1.9 TB of vectors, against about 82 GB for Tomoro's 320-dimension ColQwen3 by the same arithmetic. That gap usually decides the architecture before accuracy does.
Before you commit
- Setup
- Install transformers 4.57.2 or later and flash-attn 2.6.3, then load with trust_remote_code in bfloat16. NVIDIA tests only on Linux with A100 40GB or 80GB and H100 80GB; images go in as PIL objects.
- Integrations
- Transformers and the MTEB evaluation harness. No NIM, hosted API or Hugging Face inference provider is listed, and retrieval needs a vector database with multi-vector (MaxSim) scoring.
- Data and privacy
- Self-hosted, so documents stay on your hardware. The 4B and 8B are CC BY-NC 4.0 and the 3B is under NVIDIA's Non-Commercial License; both bar commercial use, and the 3B also inherits the Llama 3.2 licence.
- Cost at scale
- There is no paid tier. The cost is GPU time to index plus storage: by our arithmetic up to about 19 MB a page uncompressed, roughly 19 TB per million pages, against about 0.82 TB per million for Tomoro's ColQwen3 8B by its card.
- Limits
- Evaluated up to 10,240 tokens; up to eight tiles plus a thumbnail per image. The card's model links are inconsistent and the 4B link in its tables is broken.
What's good
- 63.54 average on ViDoRe v3 for the 8B, per NVIDIA
- Indexes page images directly, no OCR or layout parsing
- 3B, 4B and 8B sizes to trade accuracy for cost
What's not
- Non-commercial licence on all three sizes
- 4,096 dimensions per token: up to about 19 MB a page uncompressed
- Linux, A100 or H100, and pinned flash-attn 2.6.3 only
- No hosted endpoint or NIM listed
Best for
Researchers and evaluation teams who need a strong open baseline for visual document retrieval and do not plan to ship it.
Alternatives
Nemotron ColEmbed V2: common questions
- Is Nemotron ColEmbed V2 worth it?
- Accurate enough to be the yardstick for page-image search, and licensed so that a yardstick is all most companies can use it as. We rate it 3.8 out of 5.
- How much does Nemotron ColEmbed V2 cost?
- Free weights for non-commercial and research use (CC BY-NC 4.0).
- What are the downsides of Nemotron ColEmbed V2?
- Non-commercial licence on all three sizes; 4,096 dimensions per token: up to about 19 MB a page uncompressed; linux, A100 or H100, and pinned flash-attn 2.6.3 only; no hosted endpoint or NIM listed.
- Who is Nemotron ColEmbed V2 best for?
- Researchers and evaluation teams who need a strong open baseline for visual document retrieval and do not plan to ship it.
- How much does Nemotron ColEmbed V2 cost for a team or in production?
- There is no paid tier. The cost is GPU time to index plus storage: by our arithmetic up to about 19 MB a page uncompressed, roughly 19 TB per million pages, against about 0.82 TB per million for Tomoro's ColQwen3 8B by its card.
- Is Nemotron ColEmbed V2 safe for company data?
- Self-hosted, so documents stay on your hardware. The 4B and 8B are CC BY-NC 4.0 and the 3B is under NVIDIA's Non-Commercial License; both bar commercial use, and the 3B also inherits the Llama 3.2 licence.
- What are the best Nemotron ColEmbed V2 alternatives?
- The closest alternatives we have reviewed are Qwen3-VL-Embedding, Cohere Embed 5 and Jina Embeddings v5.
Answers checked Oct 22, 2026.