Embedding Models Cheat Sheet

What embedding models do

Embedding models convert text, code, or images into high-dimensional numerical vectors. Those vectors are the foundation of semantic search, Retrieval-Augmented Generation (RAG), and recommendation engines: things that mean the same thing end up near each other in vector space.

The frontier models to know

• Voyage 4 series (Voyage AI) — first production-grade embedding model built on a Mixture-of-Experts architecture; higher retrieval accuracy at lower serving cost.
• Nemotron 3 Embed (NVIDIA) — tops retrieval benchmarks like RTEB; an 8B version for precision-critical retrieval plus small hardware-accelerated variants for high throughput.
• Arctic Embed 2.0 (Snowflake) — open-source, multilingual, designed to run well on standard infrastructure.
• Jina Embeddings v3 (Jina AI) — ~570M parameters, multilingual; a favorite for complex RAG pipelines and cross-lingual search.
• text-embedding-3 large & small (OpenAI) — the industry-standard closed option; supports Matryoshka truncation (below).

The full list, with links to weights and docs, lives in the model repository.

How they're rated: leaderboards

• MTEB (Massive Text Embedding Benchmark) — the source of truth, hosted on Hugging Face. Aggregates 56+ datasets across 8 task categories (retrieval, clustering, classification, semantic similarity, bitext mining) into one macro-average score.
• RTEB — a stricter, precision-focused retrieval benchmark where the heavyweight models compete on raw accuracy.

How they're rated: retrieval metrics

• Recall@K / Hit Rate — how often the correct document appears in the top K results (usually top 1, 5, or 10).
• NDCG@K — measures ranking quality: it penalizes a model that puts the most relevant document at position 9 instead of position 1.
• Spearman correlation — for similarity tasks: does the mathematical distance between two vectors line up with human similarity judgments?

How they're rated: production stress tests

• Matryoshka truncation — can you shrink a vector from 3072 dimensions to 256 (to cut vector-database costs) without losing much accuracy?
• Long-context / needle-in-a-haystack — can it pull one precise fact out of a 32,000-token document without getting lost in the noise?
• Cross-lingual alignment — does a concept in one language land near the same concept in another (not just a literal translation)?

Choosing one: the three questions

1. What languages do you need?
2. Is your data text-only, or multimodal (code, images)?
3. Managed API, or open weights you can run locally?

Answer those three and the leaderboards will narrow your shortlist fast. Check current rankings on the MTEB leaderboard before committing — this space moves quickly.