ModernVBERT
ModernVBERTModernVBERT is a 250M-parameter vision-language encoder built for document retrieval rather than repurposed from a generative decoder. The paper walks the whole training pipeline - attention masking, image resolution, modality-alignment data regime, late-interaction contrastive objective - and the resulting ColModernVBERT checkpoint matches visual-document retrievers ten times its size while running query encoding on CPU. The suite also ships a bi-encoder variant, the aligned base model and the intermediate checkpoints.
From the same authors as ColPali - Teiletche, Mace, Conti, Loison, Viaud, Colombo and Faysse - and released under its own ModernVBERT Hub org rather than under vidore. The colpali-engine branch that loads it was not merged at the time of writing, which the card says outright.
Openness
5 high confidence- weights
- open(ungated MIT checkpoints on the ModernVBERT org - colmodernvbert, bimodernvbert, modernvbert-embed and the base model)
- data
- open(the paper states models, code and data are released, and the org publishes the NatCap modality-alignment corpus alongside the public document-retrieval sets it trains on)
- code
- open(the card releases the training codebase under MIT, and the paper releases intermediate checkpoints with it)
- checkpoints
- open(ModernVBERT/ablation_checkpoints, the intermediate and ablation runs behind the paper)
- license
- MIT(OSI)
MIT covers the architecture, weights and training codebase, the data is published in the same org, and the intermediate checkpoints are kept rather than discarded. The training corpus is a mixture of public sets plus the released NatCap corpus rather than one named dataset, but nothing here is withheld.
- https://arxiv.org/abs/2510.01149 recorded 2026-09-11
'Models, code and data are available at this https URL.'
- https://huggingface.co/api/datasets?author=ModernVBERT&limit=50 recorded 2026-09-11
the org publishes ModernVBERT/natcap and ModernVBERT/toy-colpali-train-set as public datasets.
- https://huggingface.co/api/models?author=ModernVBERT&limit=50 recorded 2026-09-11
seven public repos including ModernVBERT/ablation_checkpoints; `gated: false` throughout.
- https://huggingface.co/ModernVBERT/colmodernvbert/raw/main/README.md recorded 2026-09-11
`license: mit` front matter and 'We release the ModernVBERT model architectures, model weights, and training codebase under the MIT license.'; lists ColModernVBERT, BiModernVBERT, ModernVBERT-embed and the base model.
Adoption
3 high confidenceAdoption is measured as Hugging Face downloads for ColModernVBERT, which accounts for effectively the whole family since the bi-encoder, base and ablation repos draw only single digits.
- https://huggingface.co/api/models/ModernVBERT/colmodernvbert recorded 2026-09-11
`downloads` (trailing 30 days) = 132434; `gated: false`; `private: false`.
Capability
3 high confidenceModernVBERT's published numbers are credible and it leads every other sub-1B model on the same table, but the read here is about narrow scope, not dated evidence: it is English-only at 250M parameters and trails ColPali by roughly a tier on the same benchmark.
- https://arxiv.org/abs/2510.01149 recorded 2026-09-11
'a compact 250M-parameter vision-language encoder that outperforms recent models up to 10 times larger when fine-tuned on document retrieval tasks'.
- https://arxiv.org/html/2510.01149v1 recorded 2026-09-11
Table 3, ViDoRe Leaderboard, nDCG@5: ColModernVBERT (ours) 0.25B 81.2 / 56.0 / 68.6 at .032s CPU query latency; ColPali 2.92B 81.6 / 56.8 / 69.2; ColQwen2.5 3.75B 89.5 / 61.5 / 75.5; ColFlor 0.17B 68.8 / 43.0 / 55.9.
Verified 2026-09-11, except some scores, still awaiting re-confirmation