jinaai/jina-embeddings-v4

Visual Document Retrieval • 4B • Updated Apr 8 • 139k • 539

Note 2048 dimensions, $0.12 per 1M tokens. With 8k-token prompts this card prefills at roughly 170k-220k tok/s, which makes long-context retrieval the most cost-effective workload on it.

BAAI/bge-reranker-v2-m3

Text Classification • 0.6B • Updated Jun 24, 2024 • 16.9M • • 1.23k

Note $0.03 per 1M tokens. The cheapest quality improvement in a RAG pipeline: embedding search gets you 20 plausible passages, the reranker turns them into the 4 that answer the question.

Note Live demo. A full embed, retrieve, rerank and generate loop across three models on one key, with the latency and cost of each stage reported separately.