Releases·Model Release
Tencent releases EVIE-8B and EVIE-4.5B visual document retrieval models, topping ViDoRe V3 at 66.75 nDCG@10
Tencent released EVIE-8B and EVIE-4.5B on Hugging Face, models for high-capacity visual document retrieval. EVIE-8B reports state-of-the-art retrieval performance with 66.75 nDCG@10 on ViDoRe V3 and 4096D per-token multi-vector embeddings that preserve layout, typography, charts, and tables.
14 engagement1 sourceMon, Sep 7, 2026, 09:47 AM
AI Overview
- EVIE-8B scores 66.75 nDCG@10 on ViDoRe V3 with 4096D per-token multi-vector embeddings preserving layout, typography, charts, and tables
- EVIE-4.5B uses single-projection Prefix-MRL, truncating at runtime from 2048D to 64–1024D without separate models
- EVIE-4.5B scores 66.02 on ViDoRe V3, distilled from the 8B teacher via capacity-aware EVIE-ARD relation distillation
- Training-free HAC clustering compresses ~750 tokens/page to 32 vectors, cutting index storage to 3.81 GiB per million pages
- Both models evaluated across 138 multilingual tasks spanning ViDoRe V1–V3 and JinaDr under nDCG, Recall, MAP, and MRR metrics
Related
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
Releases·Model Release
Cohere launches Embed 5 Pro and Fast, embedding models with shared embedding space and RCP-nDCG@10 eval
Cohere launched Embed 5, a two-model embeddings family: Embed 5 Pro, which posts the top average score on ViDoRe V3 ahead of Voyage 4, Gemini Embedding 2, and Jina Embeddings v5, and Embed 5 Fast, a lighter low-latency model for high-volume use. Both models share a single embedding space, so teams can index with Pro and query with either without re-indexing. Embed 5 is also Cohere's first family evaluated with its new RCP-nDCG@10 retrieval methodology.
Cohere launched Embed 5, a two-model embeddings family: Embed 5 Pro, which posts the top average score on ViDoRe V3 ahead of Voyage 4, Gemini Embedding 2, and Jina Embeddings v5, and Embed 5 Fast, a lighter low-latency model for high-volume use. Both models share a single embedding space, so teams can index with Pro and query with either without re-indexing. Embed 5 is also Cohere's first family evaluated with its new RCP-nDCG@10 retrieval methodology.
2 engagement1 sourceredditWed, Sep 30, 2026, 01:20 PM
Developments·Paper
DistilVDR: 524M visual document retriever distilled from 8B teacher
Researchers propose DistilVDR, a 524M-parameter end-to-end visual document retriever distilled from a single 8B vision-language teacher using dual-student distillation and pointwise cosine alignment. It aims to reduce indexing and serving costs versus multi-billion-parameter models while maintaining retrieval quality.
Researchers propose DistilVDR, a 524M-parameter end-to-end visual document retriever distilled from a single 8B vision-language teacher using dual-student distillation and pointwise cosine alignment. It aims to reduce indexing and serving costs versus multi-billion-parameter models while maintaining retrieval quality.
0 engagement1 sourcearxivTue, Aug 11, 2026, 08:23 AM
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.
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 sourcerssThu, Sep 3, 2026, 01:13 PM
Developments·Paper
Giga-Embeddings: Sparse MoE encoder for high-throughput text embeddings
A new family of text embedding models, Giga-Embeddings, is introduced with a focus on retrieval quality and serving efficiency. The largest model is a sparse 10B-parameter Mixture-of-Experts encoder with 1.8B active parameters. It achieves top performance on MTEB benchmarks and offers high throughput, processing 114.5k tokens per second in tests.
A new family of text embedding models, Giga-Embeddings, is introduced with a focus on retrieval quality and serving efficiency. The largest model is a sparse 10B-parameter Mixture-of-Experts encoder with 1.8B active parameters. It achieves top performance on MTEB benchmarks and offers high throughput, processing 114.5k tokens per second in tests.
0 engagement1 sourcearxivMon, Aug 24, 2026, 08:16 PM