Skip to content

Benchmarks Fresh 🌱

MTEB benchmark scores and performance characteristics across different dimension settings.

MTEB Scores by Dimension

Detailed Scores

DimensionsMTEB ScoreDelta from MaxRecommended
3072Top (default)BaselineYes - highest quality
204868.16-0.01No - use 1536 instead
153668.170.00Yes - best efficiency
76867.99-0.18Yes - cost-effective
51267.55-0.62Only for very large datasets
25666.19-1.98Budget / prototype only
12863.31-4.86Minimal / experimental

Recommended Dimensions

Google recommends 3072, 1536, and 768 for highest quality. The jump from 768 to 1536 is only +0.18, while the jump from 256 to 768 is +1.80 — making 768 the sweet spot for most use cases.

Dimension Selection Guide

Key Observations

  1. 1536 matches 3072 in MTEB score (68.17) while using half the storage — a clear winner for most production systems.

  2. 768 loses only 0.18 points compared to the maximum, making it the best choice when storage or latency matters.

  3. Steep drop below 512 — the quality drops significantly from 512 (67.55) to 256 (66.19) and even more to 128 (63.31).

  4. 2048 is suboptimal — it scores 68.16, slightly below 1536 at 68.17. There is no reason to use 2048 over 1536.

Matryoshka Representation Learning (MRL)

Gemini Embedding 2 uses MRL to enable flexible dimensions. This technique "nests" information hierarchically:

  • The first 128 dimensions capture the most important features
  • Each additional block of dimensions adds finer-grained detail
  • Truncating to lower dimensions preserves the most significant information

Normalization Required

When using dimensions below 3072, normalize your vectors before computing cosine similarity. The MRL truncation preserves semantic meaning but may change vector magnitudes.

Storage and Performance Tradeoffs

DimensionsStorage/1MIndex Build TimeQuery LatencyQuality
3072~12 GBLongestHighestBest
1536~6 GBMediumMediumNear-best
768~3 GBFastLowVery good
256~1 GBFastestLowestAcceptable

Practical Recommendations

Use CaseRecommended DimWhy
Production RAG768 or 1536Best quality/cost balance
Semantic search1536Near-maximum quality
Real-time applications768Low latency
Large-scale clustering256-512Storage efficiency
Maximum accuracy3072Full quality, no tradeoffs
Prototyping768Fast iteration

See Also

Built with VitePress. Not affiliated with Google.