Choosing the right embedding model significantly impacts RAG system quality. MTEB (Massive Text Embedding Benchmark) is the standard evaluation covering retrieval, classification, clustering, and semantic similarity. Top performers (MTEB Retrieval score, 2025): 1) text-embedding-3-large (OpenAI): 54.9 nDCG@10, 3072 dims, $0.13/1M tokens - best all-around hosted option. 2) Cohere embed-v3: 55.0 nDCG@10, 1024 dims, strong multilingual, $0.10/1M tokens. 3) BGE-M3 (open source): 54.7 nDCG@10, 1024 dims, free self-hosted, supports 100+ languages and multiple retrieval methods (dense + sparse + multi-vector). 4) E5-mistral-7b (Microsoft): 56.9 nDCG@10 (SOTA), 4096 dims, 7B parameter model - expensive to run but highest quality. 5) nomic-embed-text-v1.5 (open source): competitive with OpenAI ada-002 at 8192 context length. Selection guide: Use text-embedding-3-small ($0.02/1M) for cost-sensitive applications with acceptable quality. Use BGE-M3 for self-hosted deployment with multilingual needs. Use E5-mistral for maximum quality when compute is available. Fine-tuning matters: domain-specific fine-tuning of open source models often outperforms general-purpose hosted models for specialized domains.