AI Potluck
Contribute
Back to Gap Map Model components / Embeddings & retrieval models

Nemotron Embed

NVIDIA
open weights / Overall score: 4.0(strong)

Nemotron Embed is NVIDIA's text embedding line for retrieval-augmented generation, now in its third generation. The current Nemotron-3-Embed release ships 1B and 8B bidirectional encoders with average pooling and 2048-dimension output, in BF16 and NVFP4 builds; the preceding generation - llama-embed-nemotron-8b and llama-nemotron-embed-1b-v2 - is still distributed. The line is tuned for multilingual and cross-lingual retrieval and evaluated on RTEB, MMTEB retrieval and the text split of ViDoRe-V3.

1 license text, the llama-embed-nemotron-8b LICENSE file and five Hub API records. Every checkpoint reads license: other on the Hub, so each license was read from its own body; the three generations do not share one. The visual-document (colembed, embed-vl) and Cosmos lines are separate products and are not counted here. Verified 2026-09-11 via the Nemotron-3-Embed-1B model card, the OpenMDW-1.

Openness

3 medium confidence
3.0
weights
open(ungated BF16 and NVFP4 checkpoints at 1B and 8B)
data
documented-not-released(161 dataset files over 8.5M+ points, public sources enumerated with links, synthetic generation described
code
partial(inference and pooling examples on the card
license
OpenMDW-1.1(Linux Foundation OpenMDW License Agreement v1.1, governing the current Nemotron-3-Embed release)

OpenMDW-1.1 is the Linux Foundation's Open Model, Data and Weights License Agreement v1.1, and the shared ladder places it at permissive_non_osi: it grants dealing in the model materials without restriction under copyright, patent, database and trade-secret rights, asks only that the agreement and origin notices travel with a redistribution, and explicitly disclaims any restriction on the outputs. Two complications are recorded rather than resolved: the previous generation ships under different terms, and llama-embed-nemotron-8b in particular is governed by an NVIDIA license whose section 3.3 restricts use to non-commercial research. That SKU is a superseded release rather than part of the governing one, so multi_sku_rule is not applied across the generation boundary - but anyone reaching for the 8B model from October 2025 is taking a non-commercial license.

  • https://huggingface.co/nvidia/Nemotron-3-Embed-1B-BF16/raw/main/README.md recorded 2026-09-11

    license: other with license_name openmdw-1.1 and a link to openmdw.ai; the dataset section gives 8.5M+ data points across 161 dataset files, enumerates the public datasets with links, and describes the synthetic generation, without offering the mixture for download; usage is inference and pooling code only.

  • https://openmdw.ai/license/1-1/ recorded 2026-09-11

    OpenMDW-1.1 grants permission to deal in the Model Materials without restriction under copyright, patent, database and trade secret rights, requires only that a redistribution retain the agreement and notices of origin, terminates on a patent or copyright suit, and imposes no restriction on the use, modification or sharing of outputs.

  • https://huggingface.co/nvidia/llama-embed-nemotron-8b/raw/main/LICENSE recorded 2026-09-11

    The NVIDIA License governing the previous-generation 8B checkpoint: section 3.3 states the Work and any derivative works may only be used or intended for use non-commercially, meaning non-commercial research purposes only, with NVIDIA itself excepted.

  • https://huggingface.co/api/models/nvidia/Nemotron-3-Embed-1B-BF16 recorded 2026-09-11

    gated: false; private: false; cardData license other, license_name openmdw-1.1, license_link openmdw.ai.

Adoption

4 high confidence
4.0

1,691,717 downloads in the trailing 30 days across the five text-embedding checkpoints of both live generations: 661,992 for Nemotron-3-Embed-1B-BF16, 507,552 for llama-nemotron-embed-1b-v2, 411,272 for llama-embed-nemotron-8b, 98,108 for Nemotron-3-Embed-8B-BF16 and 12,793 for the NVFP4 build. Adoption combines across the tier's releases, so the superseded generation counts here even though it does not govern openness. The visual-document and Cosmos embedding lines are excluded as separate products.

Capability

4 high confidence
4.0

Rung 4, the competitive frontier, and the placement rests on two generations agreeing. The current 1B model posts 72.38 on RTEB and 71.04 on MMTEB retrieval, and the line it replaced held the top Borda rank on MMTEB v2 in October 2025 - so this is a publisher that reaches the leaderboard rather than one claiming to be near it. One below Harrier-OSS rather than level with it because the instruments do not line up: NVIDIA reports the retrieval split and RTEB, the anchor reports the full MMTEB v2 mean of 74.3, and the two figures are not comparable. Level 5 would need the same number on the same table. Placed against the rung definition rather than a peer: no single measurement reports this product and a category peer together - RTEB and MMTEB-Retrieval against the anchor's MMTEB v2 mean - and rule (d) forbids an edge without one.

Verified 2026-09-11