--- title: NeoMME: an efficient Multimodal-native and Multilingual... description: H Company introduces NeoMME, a family of 260M and 800M multilingual multimodal encoders that process text and raw image patches in a single bidirectional... canonical: https://daily.dev/posts/neomme-an-efficient-multimodal-native-and-multilingual-encoder-oiazj2htm twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: NeoMME: an efficient Multimodal-native and Multilingual Encoder | daily.dev og:description: H Company introduces NeoMME, a family of 260M and 800M multilingual multimodal encoders that process text and raw image patches in a single bidirectional... og:url: https://daily.dev/posts/neomme-an-efficient-multimodal-native-and-multilingual-encoder-oiazj2htm og:image: https://api.daily.dev/og/posts/oIAzj2HTM.png og:image:alt: NeoMME: an efficient Multimodal-native and Multilingual Encoder og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # NeoMME: an efficient Multimodal-native and Multilingual Encoder **[Hugging Face](https://daily.dev/sources/huggingface)** · 12 min read · 0 upvotes · 0 comments ## Summary H Company introduces NeoMME, a family of 260M and 800M multilingual multimodal encoders that process text and raw image patches in a single bidirectional Transformer trained from scratch with a masked discrete-diffusion objective, rather than combining a separate vision tower with a causal language model. NeoMME-Retriever, a fine-tuned version for visual document retrieval using ColPali's page-image approach, returns dense and late-interaction embeddings in one forward pass, sits on the ViDoRe v3 Pareto frontier, and encodes about 51 pages per second on an NVIDIA L40S at 2048×2048 resolution, roughly twice ColModernVBERT's throughput. Hierarchical token pooling combined with asymmetric quantization shrinks late-interaction index storage from about 1.5 MB to 6 kB per page (255x smaller) while keeping over 95% of baseline nDCG@10. All checkpoints are released under Apache 2.0 and integrated into Hugging Face Transformers and Sentence Transformers. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much can hierarchical token pooling and asymmetric quantization shrink late-interaction embedding storage for visual document retrieval? They can reduce storage from roughly 1.5 MB to 6 kB per page, a 255x compression, while retaining more than 95% of baseline nDCG@10 retrieval quality. A less aggressive setting using pooling factor 10 with int8 queries and documents reduces storage to 39 kB per page (39x smaller) while keeping over 99% of baseline retrieval quality on ViDoRe v3. _Teams sizing multi-vector search indexes can follow embedding compression techniques like these on daily.dev._ ### How does NeoMME-Retriever-260M's encoding speed compare to ColModernVBERT for document page images? NeoMME-Retriever-260M encodes about 51 pages per second at a matched 2048x2048 image resolution on an NVIDIA L40S GPU, nearly twice ColModernVBERT's roughly 26 pages per second. Both the 260M and 800M NeoMME-Retriever models are also faster than compared models at smaller input resolutions. _Engineers comparing retrieval model throughput for indexing costs can track benchmarks like this on daily.dev._ ### How does the NeoMME-Retriever-260M model's retrieval accuracy compare to ColQwen2.5 despite having far fewer parameters? NeoMME-Retriever-260M scores 0.523 nDCG@10 on ViDoRe v3, within 0.002 of ColQwen2.5 while using about 14 times fewer parameters (260M versus roughly 3.6B). It is the highest-scoring model strictly below 800M parameters and lies on the model-size Pareto frontier alongside the 800M variant. _Anyone choosing a compact retriever for document search can weigh accuracy-per-parameter tradeoffs like these on daily.dev._ ## Similar posts on daily.dev - [Nemotron ColEmbed V2: Raising the Bar for Multimodal Retrieval with ViDoRe V3’s Top Model](https://daily.dev/posts/nemotron-colembed-v2-raising-the-bar-for-multimodal-retrieval-with-vidore-v3-s-top-model-9yhpl9gtr) · Hugging Face · 0 upvotes · 0 comments - [Run Multimodal Reasoning Agents with NVIDIA Nemotron on vLLM](https://daily.dev/posts/run-multimodal-reasoning-agents-with-nvidia-nemotron-on-vllm-vumht6rln) · vLLM · 0 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#transformers](https://daily.dev/tags/transformers) [View this post on daily.dev](https://daily.dev/posts/neomme-an-efficient-multimodal-native-and-multilingual-encoder-oiazj2htm) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"NeoMME: an efficient Multimodal-native and Multilingual Encoder","url":"https://daily.dev/posts/neomme-an-efficient-multimodal-native-and-multilingual-encoder-oiazj2htm","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/neomme-an-efficient-multimodal-native-and-multilingual-encoder-oiazj2htm"},"datePublished":"2026-09-03T13:16:48.514Z","dateModified":"2026-09-03T14:10:29.878Z","description":"H Company introduces NeoMME, a family of 260M and 800M multilingual multimodal encoders that process text and raw image patches in a single bidirectional...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/75556aa951dd7ee60fc1bfb8f8398484?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/75556aa951dd7ee60fc1bfb8f8398484?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hugging Face","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hugging Face","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/f1f55c67d81a4330acf5b90b26b0c8e1","url":"https://daily.dev/sources/huggingface"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/neomme-an-efficient-multimodal-native-and-multilingual-encoder-oiazj2htm","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,transformers","timeRequired":"PT12M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hugging Face","item":"https://daily.dev/sources/huggingface"},{"@type":"ListItem","position":3,"name":"NeoMME: an efficient Multimodal-native and Multilingual Encoder"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/neomme-an-efficient-multimodal-native-and-multilingual-encoder-oiazj2htm#faq","mainEntity":[{"@type":"Question","name":"How much can hierarchical token pooling and asymmetric quantization shrink late-interaction embedding storage for visual document retrieval?","acceptedAnswer":{"@type":"Answer","text":"They can reduce storage from roughly 1.5 MB to 6 kB per page, a 255x compression, while retaining more than 95% of baseline nDCG@10 retrieval quality. A less aggressive setting using pooling factor 10 with int8 queries and documents reduces storage to 39 kB per page (39x smaller) while keeping over 99% of baseline retrieval quality on ViDoRe v3. Teams sizing multi-vector search indexes can follow embedding compression techniques like these on daily.dev."}},{"@type":"Question","name":"How does NeoMME-Retriever-260M's encoding speed compare to ColModernVBERT for document page images?","acceptedAnswer":{"@type":"Answer","text":"NeoMME-Retriever-260M encodes about 51 pages per second at a matched 2048x2048 image resolution on an NVIDIA L40S GPU, nearly twice ColModernVBERT's roughly 26 pages per second. Both the 260M and 800M NeoMME-Retriever models are also faster than compared models at smaller input resolutions. Engineers comparing retrieval model throughput for indexing costs can track benchmarks like this on daily.dev."}},{"@type":"Question","name":"How does the NeoMME-Retriever-260M model's retrieval accuracy compare to ColQwen2.5 despite having far fewer parameters?","acceptedAnswer":{"@type":"Answer","text":"NeoMME-Retriever-260M scores 0.523 nDCG@10 on ViDoRe v3, within 0.002 of ColQwen2.5 while using about 14 times fewer parameters (260M versus roughly 3.6B). It is the highest-scoring model strictly below 800M parameters and lies on the model-size Pareto frontier alongside the 800M variant. Anyone choosing a compact retriever for document search can weigh accuracy-per-parameter tradeoffs like these on daily.dev."}}]} ```