Introducing zembed-1: The Best Text Embedding Model

Description

Read more about zembed-1 here: https://www.zeroentropy.dev/articles/introducing-zembed-1-the-worlds-best-multilingual-text-embedding-model In this video, we introduce zembed-1, a new embedding model designed to improve retrieval quality for AI systems. The talk explains how the model was trained using soft-target ranking signals derived from a re-ranker, how it improves retrieval performance compared to traditional InfoNCE training, and why it achieves over 7% higher Recall@100 compared to existing embedding models. 00:00 – zembed-1 Launch & Overview 03:00 – How Embedding Models Are Usually Trained 07:00 – New Training Method Using Z-Rank Signals 11:00 – Dataset & Training Pipeline 16:00 – Evaluation Metrics & Recall@K Results 20:00 – Performance vs Other Embedding Models 23:30 – Multilingual Performance & Cost Trade-offs 26:30 – Latency Benchmarks & Robustness Tests Learn more about us - https://www.zeroentropy.dev/