SYSTEM ONLINE

MODELgemini-3.1-pro

LATENCY142ms

THROUGHPUT4.2 tok/ms

● RAG // 10 lessons● EMBEDDINGS // 7 lessons● MODELS // 8 lessons● AGENTS // 6 lessons● PROD AI // 7 lessons● ENTERPRISE // 5 lessons● RAG // 10 lessons● EMBEDDINGS // 7 lessons● MODELS // 8 lessons● AGENTS // 6 lessons● PROD AI // 7 lessons● ENTERPRISE // 5 lessons

// LOADING

Embeddings/16 MIN/1 QUIZ

Embedding Models Compared — OpenAI, Cohere, Google, BGE, E5

When to pick which. Cost, quality, multilingual, license, hosting trade-offs. With Arabic-specific guidance most leaderboards miss.

The landscape, May 2026:

Model	Dim	Multilingual	Open weights	Notes
OpenAI `text-embedding-3-small`	1536 (truncatable)	Yes (decent)	No	Cheapest from OpenAI, matryoshka
OpenAI `text-embedding-3-large`	3072 (truncatable)	Yes (good)	No	Strongest OpenAI quality
Cohere `embed-multilingual-v3.0`	1024	Excellent — best for Arabic	No	Built explicitly for cross-lingual, supports 100+ languages
Google `text-embedding-004`	768	Good	No	Strong on Vertex / Gemini stack
Voyage `voyage-3`	1024	Good	No	Often tops MTEB English; pricier
BGE-M3	1024	Excellent (100+ langs incl. Arabic)	Yes (MIT)	Multi-functional: dense + sparse + multi-vector in one model
E5-mistral-7b-instruct	4096	Good	Yes	Strongest open dense embedder, slower
GTE-large / GTE-Qwen2	1024 / 1536	English / Multi	Yes	Strong open alternative

Decision framework

Pick proprietary (OpenAI, Cohere, Voyage, Google) when:

You need shipping speed, not infra ownership.
Your scale is below ~10M docs (the cost crossover).
You don't need data residency / on-prem.

Pick open-source (BGE-M3, E5-mistral) when:

Data privacy / compliance forbids sending to a vendor (banking, healthcare, regulated industries — IBM watsonx territory).
Cost-at-scale matters: $0.01/1M tokens × 1B docs = $10K vs. self-hosted ~$0.
You want to fine-tune on domain pairs.
Air-gapped deployments.

Arabic / multilingual reality check

Multilingual ≠ good at Arabic. Test these on your corpus:

Cohere embed-v3 multilingual — built with Arabic explicitly in scope, strong cross-lingual EN↔AR.
BGE-M3 — open, very strong Arabic, free.
OpenAI text-embedding-3-large — decent but lags Cohere on EN↔AR cross-lingual retrieval.

For Arabic specifically, also test:

Stemming / lemmatization preprocessing for sparse fallback (Arabic has heavy morphology).
Diacritic normalization at ingest (many docs include diacritics inconsistently).

Benchmarks to actually trust

MTEB (HuggingFace) — broad, English-heavy. Use as a first filter.
MIRACL — multilingual retrieval. Use for non-English serious eval.
BEIR — zero-shot retrieval across 18 datasets. Closer to real-world generalization.
Your domain eval set — the only one that decides production.

// KEY_TAKEAWAYS

What to Remember

Cohere embed-v3 and BGE-M3 are the safe bets for Arabic / multilingual.

Matryoshka truncation (OpenAI 3-large to 256d) often matches full-dim quality at fraction of cost.

Open-source crossover happens around 10M+ docs or any compliance-restricted environment.

MTEB is a starting point, not a verdict. Always run a domain eval.

// DECISION_MATRIX

When to Use

USE WHEN

Cohere embed-v3 or BGE-M3 for Arabic / multilingual.
OpenAI text-embedding-3-large with matryoshka for English at startup scale.
BGE-M3 self-hosted for compliance-heavy enterprise (IBM, banking).

AVOID WHEN

Don't pick by MTEB English rank if your workload is multilingual.
Don't self-host below 1M docs — proprietary is cheaper.

TRADE-OFFS

Self-hosted = ops cost, control, privacy.
Proprietary = velocity, but lock-in and per-token cost.

// INTERVIEW_GOTCHAS

What Trips People Up

THE TRAP

Defaulting to OpenAI because it's familiar.

GOOD ANSWER

Two finalists: Cohere embed-multilingual-v3 (best cross-lingual EN↔AR off-the-shelf, but vendor lock + cost) and BGE-M3 (open, equal-or-better Arabic, free at scale, fine-tunable). I'd start with Cohere for shipping speed, in parallel benchmark BGE-M3 with our actual eval set. If cost or compliance becomes a constraint, BGE-M3 is the migration target. Pre-process: normalize Arabic diacritics, optional stemming for sparse fallback. Add hybrid search (BM25 over Arabic root forms) since dense alone misses morphological variants.

// GLOSSARY

bge m3cohere embedmatryoshkamiraclmteb