EmbeddingGemma 2
Google DeepMind’s 740M multimodal embedding model maps text, images, audio and video into a shared 768-dimensional vector space.
At a glance
- Context window
- 8K tokens
- Max output
- —
- Text input / output · per 1M tokens
- — / —
- Knowledge cutoff
- 2025-01
Benchmark snapshot
Best result in each category
Embedding & RetrievalMTEB English v2
68.5Normal (No Tools) · #6 of 11
Benchmark details
2 results · Best published score per benchmark, grouped by category
| Benchmark · setting | Score | Rank |
|---|---|---|
| Embedding & Retrieval | ||
| MTEB English v2 Normal (No Tools) · Embedding Suites | 68.5 | 6 / 11 |
| MTEB Multilingual v2 Normal (No Tools) · Embedding Suites | 61.4 | 6 / 9 |
Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.
Back to benchmark snapshot ↑Specifications
- Provider
- Google Deep Mind
- Model type
- Embedding model
- Released
- October 6, 2026
- Total parameters
- 740M
- Architecture
- Dense
- Context window
- 8K tokens
- Knowledge cutoff
- 2025-01
- Input
- Text, Image, Audio, Video
- Output
- Embedding
- Weights
- Open weights
- Licence
- Apache 2.0
- Commercial use
- Commercial use allowed
Model analysis
EmbeddingGemma 2 is Google DeepMind’s open multimodal embedding model, released on October 6, 2026 under Apache 2.0. Built on Gemma 4, it maps text, code, images, video and audio into a shared 768-dimensional space for local retrieval, RAG, classification and clustering. Its largest changes concern code retrieval and multimodal inputs; multilingual text scores remain close to the first generation.
The full model has 740M parameters; text-only loading needs 270M
The official model card lists a 270M text module, a 170M vision encoder and a 300M audio encoder. Unused encoders can be disabled. All modalities share an 8,192-token context window. Default token costs are 280 per image, 140 per video frame and 25 per second of audio. Approximate limits of 29 images, 58 frames or 327 seconds of audio assume one modality and no accompanying text. The output is an embedding vector.
Code retrieval improves more than multilingual text
Google’s full-precision, 768-dimensional results put MTEB code v1 Mean(Task), NDCG@10 at 78.68 versus 68.76 for the first generation: a 9.92-point increase, or approximately 14.43%. MTEB multilingual v2 Mean(Task) rises only 0.21 points, from 61.15 to 61.36. MMEB v2 Image Mean(Task), Hit@1 is 57.28; VisDoc Mean(Task), NDCG@5 is 67.84; Video Mean(Task), Hit@1 is 50.67. These benchmarks use different metrics and should not be compared directly or merged into an unspecified MTEB score.
Shorter vectors save storage at a larger cost to multimodal quality
MRL supports 512-, 256- and 128-dimensional vectors, with re-normalization after truncation. At the same numerical precision, 128 dimensions use one sixth of the vector storage of 768 dimensions. However, the official MMEB v2 Overall score falls from 59.01 to 45.65, a 13.36-point drop; multilingual MTEB v2 falls from 61.36 to 57.89, a 3.47-point drop. Google recommends 128 dimensions mainly for text-only workloads. Support for 100+ languages does not imply equal quality across languages; omitting task prefixes also reduces text embedding quality.
Local multimodal retrieval still has deployment costs
The launch post reports approximately 191MB active RAM for text-only weights and 567MB for the full model with quantization on a Pixel 11 Pro. These are device- and configuration-specific measurements, not full-precision file sizes. Consider the second generation for local retrieval across media. Keeping EmbeddingGemma 300M may avoid rebuilding an existing text-only index when its quality is sufficient. Apache 2.0 permits commercial use, while self-hosting still incurs compute and storage costs. The official release materials do not specify hosted API pricing for this model.
Last reviewed 2026-10-07
EmbeddingGemma 2 vs competitors
Comparable results only · record is EmbeddingGemma 2's wins–losses
| Benchmark | EmbeddingGemma 2This model | vs EmbeddingGemma - 300MRecord 1–1 | vs Gemini Embedding 001Record 0–2 | vs Qwen3-Embedding-0.6BRecord 0–2 | vs Qwen3-Embedding-4BRecord 0–2 |
|---|---|---|---|---|---|
| MTEB English v2 | 68.5 | 69.7 | 73.3 | 70.5 | 74.6 |
| MTEB Multilingual v2 | 61.4 | 61.2 | 68.4 | 64.3 | 69.4 |
Primary sources
- EmbeddingGemma 2 technical reportblog.google
- GitHub repositorygithub.com
- Weights on Hugging Facehuggingface.co
Community discussion
Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.
Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.
No community comments yet.
