Choosing an Embedding Model in 2026: text-embedding-3, BGE, Voyage, Cohere

The 2026 Embedding Landscape

Choosing an embedding model in 2026 is more consequential than choosing an LLM. Embeddings define what your retrieval can find. They determine your storage cost. Switching them is expensive (re-index everything). The choice deserves more thought than most teams give it.

This piece compares the four families dominant in 2026: OpenAI text-embedding-3, BGE-M3 / BGE-large, Voyage-3, and Cohere embed-v4.

The Field

flowchart TB
    OAI[OpenAI text-embedding-3] --> Strength1[Strength: ecosystem, simplicity]
    BGE[BGE-M3 / BGE-large] --> Strength2[Strength: open-source, customizable]
    Voy[Voyage-3 / Voyage-Code] --> Strength3[Strength: code, domain variants]
    Coh[Cohere embed-v4] --> Strength4[Strength: multilingual, compression]

OpenAI text-embedding-3

The default for many teams. Two sizes: text-embedding-3-small (faster, cheaper) and text-embedding-3-large (higher quality, larger dimension).

• Strengths: easy to use; well-tested; integrated with OpenAI ecosystem
• Weaknesses: not the absolute strongest; no fine-tuning; no domain variants
• Dimensions: 1536 for small, 3072 for large; both support Matryoshka truncation
• Best for: general RAG, prototypes, English-first use cases

BGE-M3 / BGE-large

BAAI's open-weights embedding family. Strong performance, especially multilingual.

• Strengths: open-weights (run anywhere); multilingual; supports dense + sparse + multi-vector
• Weaknesses: more complex to deploy; you own the ops
• Dimensions: 1024
• Best for: multilingual RAG, on-prem deployments, fine-tuning needs

Voyage-3 / Voyage-Code

Voyage AI's family. Strong code-specific variant.

• Strengths: best open-API code embedding in 2026; domain variants for finance and medical
• Weaknesses: less mainstream than OpenAI; smaller community
• Dimensions: 1024 typical
• Best for: code RAG, domain-specific RAG (finance, medical)

Cohere embed-v4

Cohere's flagship. Strong multilingual; compression-friendly.

• Strengths: 100+ languages; Matryoshka support; binary / int8 quantization
• Weaknesses: less common in tooling than OpenAI
• Dimensions: configurable (256, 512, 1024)
• Best for: multilingual, latency-sensitive, storage-sensitive

Decision Matrix

flowchart TD
    Q1{Multilingual?} -->|Yes| Q2{Open-source needed?}
    Q1 -->|No, English-first| Q3{Code or domain-specific?}
    Q2 -->|Yes| BGE2[BGE-M3]
    Q2 -->|No| Coh2[Cohere embed-v4]
    Q3 -->|Yes| Voy2[Voyage]
    Q3 -->|No, general purpose| OAI2[text-embedding-3-large]

What Matters in Practice

Beyond raw recall numbers, three dimensions decide:

• Domain fit: does the model handle your vocabulary well?
• Operational fit: hosted API or self-hosted?
• Cost fit: per-million-token cost adds up at scale

Recall Benchmarks

April 2026 numbers (varies by domain, your mileage will vary):

Model	MTEB v2 average	Code (CodeSearchNet)	Multilingual (MIRACL)
OpenAI 3-large	75	71	65
BGE-M3	73	70	70
Voyage-3	76	77	67
Cohere embed-v4	74	70	73

These shift release-to-release. Run your own benchmark.

Storage Cost

Embedding dimensions affect storage:

• 3072-dim float32: ~12 KB/vector
• 1024-dim float32: ~4 KB/vector
• 1024-dim int8 (quantized): ~1 KB/vector
• Binary embeddings: ~128 bytes/vector

For a 10M-vector corpus, the difference between 3072-dim float32 and quantized 1024-dim is 100GB+ — real money.

Matryoshka Embeddings

Some models (OpenAI, Cohere) support Matryoshka: the same vector can be truncated to lower dimensions while preserving most of the quality. Useful for storage / latency optimization without re-embedding.

Migration Strategy

Switching embedding models is painful:

• Re-embed the entire corpus
• Update queries to use new model
• Cannot mix old and new embeddings (distances are not comparable)

Plan for it: tag every embedding with model version; have a re-embed pipeline ready; do migrations during low-traffic windows.

What Production Teams Actually Use

In CallSphere we use text-embedding-3-small as the default for the marketing site's blog dedup (huge corpus, quality is good enough, OpenAI ecosystem fit). For domain-specific products, we use Voyage variants. For multilingual, Cohere. The picks fit the workload.

Sources

• MTEB leaderboard — https://huggingface.co/spaces/mteb/leaderboard
• OpenAI embedding documentation — https://platform.openai.com/docs/guides/embeddings
• BGE-M3 paper — https://arxiv.org/abs/2402.03216
• Voyage AI — https://docs.voyageai.com
• Cohere embed — https://docs.cohere.com/docs/embeddings