新聞

單一模型，四種模態：文字、圖像、音訊、影片。同級最佳的 1.6B 與 0.9B 全能型向量模型。

We introduce frozen-encoder model composition, a novel approach to multimodal embedding models. We build on the VLM-style architecture, in which non-text encoders are adapted to produce input for a language model, which in turn generates embeddings for all varieties of input. The backbone text embedding models and the added non-text media encoders remain frozen. We only trained the connecting components, representing 0.35% of the total weights. The resulting jina-embeddings-v5-omni suite encodes text, image, audio, and video into a single semantic embedding space, producing competitive results with models 5-7x its size.

將任何多模態大模型轉換為小型音訊向量模型，僅需 25 分之一的數據量，表現即可超越 CLAP。

一個透過讀取原始數值來對向量模型進行指紋識別的微型 Transformer。無需特徵工程。

兩款效能領先的 1B 以下多語言向量模型，現已於 Elastic Inference Service、Llama.cpp 與 MLX 上架。

Text embedding models are widely used for semantic similarity tasks, including information retrieval, clustering, and classification. General-purpose models are typically trained with single- or multi-stage processes using contrastive loss functions. We introduce a novel training regimen that combines model distillation techniques with task-specific contrastive loss to produce compact, high-performance embedding models. Our findings suggest that this approach is more effective for training small models than purely contrastive or distillation-based training paradigms alone. Benchmark scores for the resulting models, jina-embeddings-v5-text-small and jina-embeddings-v5-text-nano, exceed or match the state-of-the-art for models of similar size. jina-embeddings-v5-text models additionally support long texts (up to 32k tokens) in many languages, and generate embeddings that remain robust under truncation and binary quantization. Model weights are publicly available, hopefully inspiring further advances in embedding model development.

We present a compression method for unit-norm embeddings that achieves 1.5x compression, 25% better than the best prior lossless method. The method exploits that spherical coordinates of high-dimensional unit vectors concentrate around pi/2, causing IEEE 754 exponents to collapse to a single value and high-order mantissa bits to become predictable, enabling entropy coding of both. Reconstruction error is below 1e-7, under float32 machine epsilon. Evaluation across 26 configurations spanning text, image, and multi-vector embeddings confirms consistent improvement. The method requires no training.

Vision encoders have remained comparatively small while language models scaled from billions to hundreds of billions of parameters. This survey analyzes vision encoders across 70+ vision-language models from 2023–2025 and finds that training methodology matters more than encoder size: improvements in loss functions, data curation, and feature objectives yield larger gains than scaling by an order of magnitude. Native resolution handling improves document understanding, and multi-encoder fusion captures complementary features no single encoder provides. We organize encoders into contrastive, self-supervised, and LLM-aligned families, providing a taxonomy and practical selection guidance for encoder design and deployment.

全新 2B 視覺語言模型在多語言 VQA 上實現 SOTA，在純文字任務上沒有災難性遺忘。

We present jina-vlm, a 2.4B parameter vision-language model that achieves state-of-the-art multilingual visual question answering among open 2B-scale VLMs. The model couples a SigLIP2 vision encoder with a Qwen3 language backbone through an attention-pooling connector that enables token-efficient processing of arbitrary-resolution images. Across standard VQA benchmarks and multilingual evaluations, jina-vlm achieves leading results while preserving competitive text-only performance. Model weights and code are publicly released.