# Google's EmbeddingGemma 2 claim tops every sub-1.7B model on MTEB Code

> Google's 78.68 would beat all 33 measured models under 1.7B parameters, but 9 larger ones still score higher on the public board.

Tec N Trend data brief · 2026-10-07 · https://tecntrend.com/briefs/2026-10-07/embedding-gemma

Google released EmbeddingGemma 2 on 6 October. The Decoder reports 740 million parameters (270 million for text only), open weights, and a score of 78.68 on the MTEB code benchmark, against 68.76 for the first EmbeddingGemma. Google says it beats rivals up to twice its size on multimodal embedding tests.

The public MTEB leaderboard has no entry for the new model yet, but it does confirm the old one: EmbeddingGemma-300m scores 68.76 on MTEB (Code, v1). Of 47 measured models with a known size, the best under 1.7B parameters is F2LLM-v2-0.6B at 77.41. Google's claim would add about 1.3 points to that, still short of 9 larger models, led by 14B and 7B systems near 80.7.

Treat 78.68 as unverified until the model is evaluated on the public board with the same task set. The multimodal claim is a separate test the code benchmark does not cover.

## Key numbers
- EmbeddingGemma 2, claimed MTEB Code: 78.68
- EmbeddingGemma-300m, measured: 68.76
- Best measured model under 1.7B params: 77.41 (F2LLM-v2-0.6B)
- Measured models above 78.68: 9, all 1.7B or larger

## Chart: MTEB (Code, v1) mean score vs model size
Data: https://tecntrend.com/briefs/2026-10-07/embedding-gemma.json

## Sources
- [MTEB leaderboard API, benchmark "MTEB(Code, v1)" scores](https://mteb-leaderboard-backend.hf.space/v1/benchmarks/MTEB(Code,%20v1)/scores) (MTEB results are CC0 (mteb/results dataset), captured 2026-10-07)
- [The Decoder, EmbeddingGemma 2 report](https://the-decoder.com/google-claims-embeddinggemma-2-outperforms-rival-embedding-models-twice-its-size/) (news report (facts only), captured 2026-10-07)
- News: [Google claims EmbeddingGemma 2 outperforms rival embedding models twice its size](https://the-decoder.com/google-claims-embeddinggemma-2-outperforms-rival-embedding-models-twice-its-size/) (The Decoder, 2026-10-06)

## Caveats
- The EmbeddingGemma 2 point (740M parameters, 78.68) is Google's own claim as reported by The Decoder, not a leaderboard result. It is flagged and was never verified by us.
- The match between the leaderboard's 68.76 for EmbeddingGemma-300m and the figure The Decoder quotes for the predecessor suggests the claim uses MTEB (Code, v1), but the article does not name the benchmark version.
- Only the 47 models with a full 12-task score and a known parameter count are plotted; API-only models (no size) and models with missing tasks are left out.
- Parameter counts are the leaderboard's totals, which include embedding tables; the "270M text-only" figure for EmbeddingGemma 2 would move its x position left.
- Mean task score is a plain average over 12 code tasks; labs can train on benchmark data, and the leaderboard marks that case separately (not applied here).