Specialist modelOpen weights
EmbeddingGemma 2
Google DeepMind
- Released
- October 6, 2026
- Data date
- October 7, 2026
EmbeddingGemma 2 extends Google's EmbeddingGemma with shared representations for text, code, images, video, and audio. Its predecessor handled text; the new version maps several modalities into one 768-dimensional vector space.
Google releases 740 million parameters under Apache 2.0 for local use. The model returns embedding vectors rather than text or audio responses.
Specifications and access
| Specification | Value and source |
|---|---|
| Model ID | google/embeddinggemma-2Source |
| Input modalities | Text including code, images, video, and audio, including mixed inputsSource |
| Output modalities | 768-dimensional embedding vectors, truncatable to 512, 256, or 128 dimensions; no text or audio responsesSource |
| Parameters | 740 million total, including modular vision and audio encodersSource |
| Context window | 8,192 tokens across supported modalitiesSource |
| License | Apache 2.0Source |
| Access | Public weights on Hugging Face and Kaggle; local executionSource |
Page 1 of 3
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
MTEB multilingual v2 (output 768d) (Show measurement, test conditions, and source)
- Source value
- 61.36
- Score
- 61.36
- Metric
- Mean(Task); Multiple
- Unit
- No unit provided
- Benchmark version
- MTEB v2
- Category
- multilingual
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 768 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB multilingual v2 (output 512d) (Show measurement, test conditions, and source)
- Source value
- 61.17
- Score
- 61.17
- Metric
- Mean(Task); Multiple
- Unit
- No unit provided
- Benchmark version
- MTEB v2
- Category
- multilingual
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 512 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB multilingual v2 (output 256d) (Show measurement, test conditions, and source)
- Source value
- 60.41
- Score
- 60.41
- Metric
- Mean(Task); Multiple
- Unit
- No unit provided
- Benchmark version
- MTEB v2
- Category
- multilingual
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 256 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB multilingual v2 (output 128d) (Show measurement, test conditions, and source)
- Source value
- 57.89
- Score
- 57.89
- Metric
- Mean(Task); Multiple
- Unit
- No unit provided
- Benchmark version
- MTEB v2
- Category
- multilingual
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 128 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB English v2 (output 768d) (Show measurement, test conditions, and source)
- Source value
- 68.46
- Score
- 68.46
- Metric
- Mean(Task); Multiple
- Unit
- No unit provided
- Benchmark version
- MTEB v2
- Category
- English text
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 768 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB English v2 (output 512d) (Show measurement, test conditions, and source)
- Source value
- 68.41
- Score
- 68.41
- Metric
- Mean(Task); Multiple
- Unit
- No unit provided
- Benchmark version
- MTEB v2
- Category
- English text
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 512 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB English v2 (output 256d) (Show measurement, test conditions, and source)
- Source value
- 67.78
- Score
- 67.78
- Metric
- Mean(Task); Multiple
- Unit
- No unit provided
- Benchmark version
- MTEB v2
- Category
- English text
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 256 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB English v2 (output 128d) (Show measurement, test conditions, and source)
- Source value
- 65.68
- Score
- 65.68
- Metric
- Mean(Task); Multiple
- Unit
- No unit provided
- Benchmark version
- MTEB v2
- Category
- English text
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 128 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB code v1 (output 768d) (Show measurement, test conditions, and source)
- Source value
- 78.68
- Score
- 78.68
- Metric
- Mean(Task); NDCG@10
- Unit
- No unit provided
- Benchmark version
- MTEB v1
- Category
- code
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 768 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB code v1 (output 512d) (Show measurement, test conditions, and source)
- Source value
- 77.24
- Score
- 77.24
- Metric
- Mean(Task); NDCG@10
- Unit
- No unit provided
- Benchmark version
- MTEB v1
- Category
- code
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 512 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB code v1 (output 256d) (Show measurement, test conditions, and source)
- Source value
- 76.18
- Score
- 76.18
- Metric
- Mean(Task); NDCG@10
- Unit
- No unit provided
- Benchmark version
- MTEB v1
- Category
- code
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 256 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
MTEB code v1 (output 128d) (Show measurement, test conditions, and source)
- Source value
- 71.41
- Score
- 71.41
- Metric
- Mean(Task); NDCG@10
- Unit
- No unit provided
- Benchmark version
- MTEB v1
- Category
- code
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Google DeepMind EmbeddingGemma 2 model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Full-precision google/embeddinggemma-2 checkpoint; MRL output vector with 128 dimensions. The card specifies re-normalization after truncation; it does not identify the exact prompt prefix or task list for this run.
- Context
- The model card publishes aggregate task scores without a common unit or per-run sample count.
Page 1 of 3
Sources and data date
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 7, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | per-suite prompts, task lists, and sample counts Original wording Google DeepMind’s official card publishes aggregate results for the exact checkpoint, including the full-precision and truncated-vector tables, but not the task-level scores, prompt assignment, or sample count behind those aggregates. The official MTEB model registry has an entry for google/embeddinggemma-300m but not this exact checkpoint; its filtered results endpoint returned HTTP 500 and remained unavailable. |
| Additional source | Google EmbeddingGemma 2 announcement (retrieved October 7, 2026) · Google EmbeddingGemma 2 announcement · Editorial description reviewed October 7, 2026 |
| Additional source | Google EmbeddingGemma 2 model card (retrieved October 7, 2026) · Google EmbeddingGemma 2 model card · Editorial description reviewed October 7, 2026 |
| Additional source | Model metadata (retrieved October 7, 2026) |