Releases·Model Release
NeoMME: 260M–800M multimodal encoders trained from scratch with masked diffusion for document retrieval
A new post introduces NeoMME, a family of 260M and 800M multilingual multimodal encoders. A single bidirectional Transformer processes both text tokens and raw image patches, trained from scratch with a masked discrete-diffusion objective, avoiding a separate vision tower or causal LM. It was fine-tuned for visual document retrieval using ColPali's page-image approach.
0 engagement1 sourceThu, Sep 3, 2026, 01:13 PM
AI Overview
- NeoMME (260M/800M) uses a single bidirectional Transformer for text and raw image patches, trained from scratch with a masked discrete-diffusion objective on 524B tokens.
- NeoMME-Retriever emits dense and 128-dim late-interaction embeddings in one pass; the 260M model tops ViDoRe v3 nDCG@10 (0.523) among models under 800M.
- Hierarchical token pooling plus int8/binary asymmetric quantization shrinks late-interaction storage from 1.5 MB to 6 kB per page (255×) at >95% of baseline nDCG@10.
- On one NVIDIA L40S at 2048×2048 input, NeoMME-Retriever-260M encodes ~51 pages/second, about twice ColModernVBERT's 26.
- Checkpoints ship in Hugging Face Transformers under Apache 2.0 with a 131k-token BPE vocabulary and 16,384-token context.
Related
Releases·Model Releasefeatured
Google DeepMind releases EmbeddingGemma 2, a 740M-parameter multimodal embedding model, on Hugging Face
Google DeepMind released EmbeddingGemma 2, an Apache 2.0 open multimodal embedding model that maps text (including code), images, video, and audio into a shared 768-dimensional vector space. The 740M-parameter model runs locally on as little as 0.5GB RAM, targets on-device search, RAG, classification, and clustering, and shipped with same-day support in transformers v5.19.0 plus a WebGPU in-browser demo.
Google DeepMind released EmbeddingGemma 2, an Apache 2.0 open multimodal embedding model that maps text (including code), images, video, and audio into a shared 768-dimensional vector space. The 740M-parameter model runs locally on as little as 0.5GB RAM, targets on-device search, RAG, classification, and clustering, and shipped with same-day support in transformers v5.19.0 plus a WebGPU in-browser demo.
52010.6k engagement9 sourcesreddit, github, xTue, Oct 6, 2026, 03:41 PM
Releases·Model Release
Tencent releases EVIE-Preview-4.5B, a multilingual visual document retrieval model
Tencent released EVIE-Preview-4.5B, a multilingual visual document retrieval model built on Qwen3.5-4B. It uses ColBERT-style late interaction with 128-dimensional multi-vector embeddings, cutting vector storage costs by 8× to 32×. The model achieves top-tier performance on ViDoRe benchmarks.
Tencent released EVIE-Preview-4.5B, a multilingual visual document retrieval model built on Qwen3.5-4B. It uses ColBERT-style late interaction with 128-dimensional multi-vector embeddings, cutting vector storage costs by 8× to 32×. The model achieves top-tier performance on ViDoRe benchmarks.
18 engagement2 sourcesreddit, rssMon, Aug 17, 2026, 12:24 PM
Developments·Paper
Decoupled Visual Processing: Efficient Multimodal Adaptation via Modality-Specific Transformer Substitution
Researchers propose Decoupled Visual Processing (DVP), a training framework that replaces upper decoder layers of a pretrained LLM with lightweight modality-specific modules for visual instruction tuning. This approach reduces computational cost by avoiding full fine-tuning of all parameters, leveraging the divergence in representation requirements between visual and textual tokens in deeper layers.
Researchers propose Decoupled Visual Processing (DVP), a training framework that replaces upper decoder layers of a pretrained LLM with lightweight modality-specific modules for visual instruction tuning. This approach reduces computational cost by avoiding full fine-tuning of all parameters, leveraging the divergence in representation requirements between visual and textual tokens in deeper layers.
0 engagement1 sourcearxivWed, Jul 29, 2026, 08:16 AM
Releases·Model Release
CohereLabs releases North Micro Vision Instruct, a 2.4B open-weight VLM
CohereLabs released North Micro Vision Instruct, a 2.4B-parameter open-weight vision-language model under Apache 2.0. It supports native-resolution images, multilingual and multi-image inputs, and is designed for prototyping and fine-tuning.
CohereLabs released North Micro Vision Instruct, a 2.4B-parameter open-weight vision-language model under Apache 2.0. It supports native-resolution images, multilingual and multi-image inputs, and is designed for prototyping and fine-tuning.
122 engagement1 sourceredditWed, Aug 12, 2026, 04:50 PM
Developments·Paper
MIDR paper proposes training-free enrichment-augmented indexing for multimodal document retrieval
A new arXiv paper introduces MIDR (Multimodal Indexing for Document Retrieval), a training-free framework that moves multimodal reasoning from query time to index time for retrieval over visually rich documents. It targets the weakness of plain OCR pipelines, which linearize or drop content in tables, charts, figures, and layout relations. It offers an alternative to ColPali-family visual retrievers that keep image-derived retrieval on the query-time serving path.
A new arXiv paper introduces MIDR (Multimodal Indexing for Document Retrieval), a training-free framework that moves multimodal reasoning from query time to index time for retrieval over visually rich documents. It targets the weakness of plain OCR pipelines, which linearize or drop content in tables, charts, figures, and layout relations. It offers an alternative to ColPali-family visual retrievers that keep image-derived retrieval on the query-time serving path.
0 engagement1 sourcearxivTue, Sep 1, 2026, 02:38 PM