Report
NeoMME ships a 260M and 800M multimodal encoder, with a retriever the team says cuts late-interaction index storage 255×
H Company has released NeoMME, a family of 260M and 800M multilingual multimodal encoders trained from scratch without a separate vision tower or a causal language model. Its retrieval variant returns dense and late-interaction embeddings in one forward pass, and the team reports large reductions in index storage.
Most visual document retrievers in circulation are adapted from generative vision-language models: a pretrained vision encoder produces visual features, a projector maps them into a language model's input space, and a causal decoder processes the combined representation. NeoMME takes a different route. According to the release post, a single bidirectional Transformer processes both text tokens and raw image patches, and the entire model is trained from scratch with a masked discrete-diffusion objective. Images are split into non-overlapping 32×32 patches projected with a small MLP; text uses factorized token embeddings over a 131k-token BPE vocabulary built from multilingual text, code, mathematics and machine-produced image transcripts.
Both variants share one architecture and a context length of 16,384 tokens, which the post says is enough for up to two standard 3840×2160 4K UHD images. Images keep their aspect ratio and size, so high-resolution document pages consume more tokens than low-content images. Most layers use symmetric sliding-window attention, with global attention every sixth layer and in the final layer.
The retrieval head is the practical part. NeoMME-Retriever was fine-tuned for visual document retrieval following ColPali's page-image approach and returns dense and late-interaction embeddings in a single forward pass. The team reports both model sizes on the ViDoRe v3 Pareto frontier for nDCG@10 and model size, and says that at a matched 2048×2048 image input size on an NVIDIA L40S GPU the 260M model encodes about 51 pages per second, roughly twice ColModernVBERT's throughput.
The storage figure is the one to test against your own setup. Hierarchical token pooling and asymmetric quantization are said to reduce late-interaction index storage from roughly 1.5 MB to 6 kB per page — 255× smaller — while retaining more than 95% of baseline nDCG@10. NeoMME is available in Hugging Face Transformers, and all checkpoints are released under the Apache 2.0 license.
Our reading
For teams whose visual document retrieval cost is dominated by index storage or GPU seconds rather than by leaderboard position, the storage and throughput claims are the ones worth reproducing, because they change what fits on a single machine. The people who should care are those indexing scanned or rendered page images and maintaining late-interaction indexes, particularly in multilingual corp…
What to do or watch
Run a sample of your own page images through NeoMME-Retriever at the same resolution you currently use and compare nDCG@10 and index size against your existing retriever before switching anything. Whether the 95% nDCG@10 retention holds outside the team's own evaluation is not established by the evidence available here.
Source details and supporting facts
Each line is stated by the page named above it.
Stated by Hugging Face
- NeoMME is introduced as a family of 260M and 800M multilingual multimodal encoders that use a single bidirectional Transformer rather than a separate pretrained vision tower or a causal language model.
- The 260M model encodes about 51 pages per second at a matched 2048×2048 image input size on an NVIDIA L40S GPU, about twice ColModernVBERT's throughput.
- Hierarchical token pooling and asymmetric quantization reduce late-interaction index storage from roughly 1.5 MB to 6 kB per page, 255× smaller, while retaining more than 95% of baseline nDCG@10.
- Both NeoMME variants have a context length of 16,384 tokens, described as enough for up to two standard 3840×2160 4K UHD images.
- NeoMME-Retriever was fine-tuned for visual document retrieval using ColPali's page-image approach and returns dense and late-interaction embeddings in one forward pass.
- NeoMME is available in Hugging Face Transformers and all model checkpoints are released under the Apache 2.0 license.
Sources
- Hugging FaceText stored 16 September 2026
How this story was checked. Written from the 1 page listed above, stored 16 September 2026; claims checked against that stored text on 16 September 2026.
What that means
- 6 of 6 reported statements were confirmed against the page that carries them; the rest were removed rather than published.
- Figures in the text were required to appear in the stored source text: yes. Identifiers: yes.
- The check reads stored text only: no claim rests on a fresh look that did not happen.
- Where the reporting was silent, the text says so instead of filling the gap.