768-dimension CLS embeddings
According to the model card, the Sentence Transformer outputs 768 dimensions with CLS pooling, normalization, and cosine similarity over up to 512 tokens.
Open Source Model Profile · qihoo360
Zhinao-ChineseModernBert-Embedding
Zhinao-ChineseModernBert-Embedding is a 227M-parameter Chinese semantic-embedding model from qihoo360. According to the model card, it supports 512-token inputs with 768-dimension vectors for retrieval and similarity workloads.
- Publisher
- qihoo360
- Task
- sentence-similarity
- Model type
- modernbert
- License
- apache-2.0
- Library
- sentence-transformers
- Publication status
- Accepted · not indexed
Model overview
Zhinao-ChineseModernBert-Embedding is published by qihoo360 as a sentence-similarity embedding model. Captured Safetensors metadata reports 227,018,496 parameters, about 227M. According to the model card, it is the specialized embedding member of the Zhinao-ChineseModernBert series, optimized for semantic retrieval, vector databases, and RAG augmentation on top of a ModernBert base.
Recorded capabilities
ModernBert plus Qwen2 tokenizer
According to the model card, the series uses the ModernBert architecture with a Qwen2Tokenizer to reduce out-of-vocabulary rates on Chinese and mixed text.
Two-stage embedding training
According to the model card, training runs retrieval contrastive pre-training followed by multi-task fine-tuning on CMTEB and MTEB task data.
Publisher-reported 1T-token base
According to the model card, the base was pre-trained from scratch on 1T high-quality Chinese and English tokens with over 65% Chinese content.
Use cases in the source record
- Chinese semantic retrieval and vector-database indexing using the documented 768-dimension normalized embeddings.
- RAG retrieval augmentation and text similarity, clustering, and reranking flows described in the model card.
- Instruction-prompted encoding with the card's documented retrieval prompts, such as the Chinese search-query instruction.
Limitations and unknowns
- No independently measured evaluation results were extracted; benchmark leadership statements are publisher-reported card claims.
- No VRAM, hardware, or pricing detail was extracted from this record.
- The model card's performance comparisons against larger models come from the publisher and were not independently verified.
- Provider state is historical snapshot data, not independently refreshed current availability.
Source and provenance
Source: qihoo360/Zhinao-ChineseModernBert-Embedding
Captured: Unknown. Processed: 2026-09-07T19:36:00.702684+00:00.
Zhinao-ChineseModernBert: 面向高吞吐低内存场景的中文基座与向量嵌入模型 项目简介 Zhinao-ChineseModernBert系列是针对 高推理速度要求、严苛内存限制 的工业级场景，从头预训练的中文Base级基座模型与语义嵌入模型。本系列基于ModernBert高效架构与Qwen2Tokenizer分词器，依托超大规模中英文语料完成全流程预训练，在保持Base级参数量（除Embedding外约100M参数）轻量化优势的同时，实现了对同量级模型的全面超越，甚至性能优于更大参数量的主流模型，为中文NLP理解任务、语义检索、向量数据库、RAG检索增强等场景提供高性价比的开箱即用解决方案。 本项目包含两个核心模型： Zhinao-ChineseModernBert ：通用中文理解基座，基于两阶段掩码语言建模(MLM)预训练，支持最长1536序列长度，适配各类中文NLU下游任务。 Zhinao-ChineseModernBert-Embedding ：专业中文语义嵌入模型，在基座的基础上做两阶段Embedding训练，支持最长512序列长度，专为语义检索、向量表征、相似度计算等场景深度优化。 核心亮点 1. 高效架构+先进分词体系，兼顾速度与泛化性 采用 ModernBert 高效Transformer架构从头预训练，针对高吞吐推理、长文本处理场景做了深度优化，相比传统Bert架构实现显著的推理速度提升，内存占用更友好。 适配 Qwen2T…
F001F002F003F004F005F006F007F010F011F012F013F014F015F016