Jina AI(@JinaAI_)2026年2月19日

v5-text uses decoder-only backbones with last-token pooling instead of mean pooling. Four lightweigh...

7.5Score

用这条生成生成视频方案

v5-text uses decoder-only backbones with last-token pooling instead of mean pooling. Four lightweigh...

AI 深度提炼

v5-text 使用 decoder-only 主干网络和 last-token pooling 替代传统 mean pooling
每个 Transformer 层注入四个轻量 LoRA 适配器，分别处理检索、匹配、分类和聚类
支持 32K 上下文长度，是 v3 的 4 倍，并在推理时动态选择适配器

#LoRA#大模型#文本嵌入#Jina AI#Transformer

打开原文

Jina AI on X: "v5-text uses decoder-only backbones with last-token pooling instead of mean pooling. Four lightweight LoRA adapters are injected at each transformer layer, handling retrieval, text-matching, classification, and clustering independently. Users select the appropriate adapter at https://t.co/j4RESCgE7S" / X

Don’t miss what’s happening

People on X are the first to know.

Post

See new posts

Conversation

![Image 2](http://x.com/JinaAI_)

Jina AI

@JinaAI_

v5-text uses decoder-only backbones with last-token pooling instead of mean pooling. Four lightweight LoRA adapters are injected at each transformer layer, handling retrieval, text-matching, classification, and clustering independently. Users select the appropriate adapter at inference time. For retrieval, queries get a "Query:" prefix and documents get "Document:". Context length is 32K tokens, a 4x increase over v3.

![Image 3: Image](http://x.com/JinaAI_/status/2024505349181755760/photo/1)

3:24 PM · Feb 19, 2026

2,570 Views