Research note
Research note: Google releases EmbeddingGemma 2, an open model that maps text, images, audio and video into one search space
How we checked the specs of EmbeddingGemma 2 against Google’s two official posts, and which figures are Google’s own claims.
What we set out to check
The model’s size, licence, input length, output sizes, memory use on phones and benchmark claims.
What we found
- Both official posts agree on 740 million parameters, the 270 million text-and-code subset, 8,192 tokens, 768 dimensions and Apache 2.0.
- Memory and benchmark numbers come only from Google.
Numbers that differ
The announcement gives a code score rising from 68.76 to 78.68 (9.92 points). The developer guide calls it 14% higher. Both are consistent: 9.92 is about 14% of 68.76.
Labels we chose and why
- Confirmed for specs Google states officially.
- Company claim for memory use and benchmark scores, which no independent tester had published when we checked.
What we could not verify
- Independent benchmark results.
- The list of supported languages, which the posts do not give.
What we searched
- EmbeddingGemma 2 Google developers blog
- EmbeddingGemma 2 Apache 2.0 740M parameters
- EmbeddingGemma 2 MTEB code benchmark
Every source we read
| Source | Used? | Why |
|---|---|---|
| Google: EmbeddingGemma 2 announcement | USED | Official announcement. Used for specs, memory figures, benchmarks and availability. |
| Google Developers Blog: developer guide | USED | Official guide. Confirms sizes, dimensions, context and licence, and gives the 14% code gain. |
| Third-party benchmark write-ups | NOT USED | Repeats Google’s numbers without independent testing. |