{
  "id": 5338693,
  "title": "NeoMME: an efficient Multimodal-native and Multilingual Encoder",
  "url": "https://urgent.news/2026/09/03/neomme-an-efficient-multimodal-native-and-multilingual-encoder",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-03T13:13:48.000Z",
  "source": {
    "name": "Hugging Face",
    "slug": "hugging-face",
    "url": "https://huggingface.co/blog/Hcompany/neomme"
  },
  "original_language": "en",
  "account": "NeoMME is a new family of 260M and 800M multilingual multimodal encoders designed to efficiently process both text and images in a single Transformer encoder. Unlike many generative visual language models that rely on separate pretrained vision towers or causal language models, NeoMME handles text and images using the same computational path. This allows for easier pretraining, fine-tuning, parallelization, and serving for both modalities.\n\nNeoMME uses a masked discrete-diffusion objective for training, which enables it to generate vector representations for input text and/or images. For text-only examples, the model samples a corruption rate between 0 and 1, independently masks each eligible text token at that rate, and reconstructs the masked text. For multimodal examples, the corruption rate ranges from 0.3 to 1, with image patches remaining visible during reconstruction.\n\nThe 260M model processes approximately 51 pages per second on an NVIDIA L40S GPU, nearly twice the throughput of ColModernVBERT. To improve storage efficiency, hierarchical token pooling and asymmetric quantization reduce late-interaction index storage from about 1.5 MB to just 6 kB per page, while retaining over 95% of baseline nDCG@10 performance.\n\nNeoMME is available in Hugging Face Transformers and all model checkpoints are released under the Apache 2.0 license. The model has demonstrated superior performance in visual document retrieval using the page-image methodology introduced by ColPali. NeoMME-Retriever, which expands on the NeoMME backbone, ranks document page screenshots and bypasses preprocessing OCR steps, preserving layout and visual clues that may not be captured by perfect OCR models.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}