EmbeddingGemma 2
Google DeepMind's 740M Apache 2.0 embedding model that puts text, code, images, video and audio in one search space; runs on-device.
In short. An embedding model turns text, pictures, video or sound into a list of numbers, so that things with similar meaning end up close together. That is what powers semantic search and retrieval-augmented generation. EmbeddingGemma 2 does this for all of those kinds of content in one shared space, so you can search video with a voice memo or find photos with a sentence. It is small enough to run on a phone or laptop, and the license is Apache 2.0.
Summary
Google DeepMind released EmbeddingGemma 2 on 6 October 2026 under the Apache 2.0 license; the first EmbeddingGemma used the Gemma license. It has 740 million parameters in total and is built on the Gemma 4 architecture. It maps text (including source code), images, video and audio, alone or in combination, into a single 768-dimensional vector space, understands more than 100 languages and has an 8,192-token context window. Google says that is enough for about 5.5 minutes of audio, 29 images or 58 video frames.
The design is modular. A 270-million-parameter text and code model is the core, and you can load a 170-million-parameter vision encoder and a 300-million-parameter audio encoder only if you need them, giving 270M, 440M (text and vision), 570M (text and audio) or 740M parameters. Matryoshka training lets you shorten the vectors from 768 to 512, 256 or 128 dimensions, which cuts vector-database storage by up to six times; Google reports that quality holds up well down to 256 dimensions. Text inputs work best with short task prefixes such as a search-query or document prefix. Google says that, with quantization, the text-only model needs about 191 MB of RAM and the full multimodal model about 567 MB on a Pixel 11 Pro phone. The BF16 checkpoint is about 1.5 GB.
Google reports a score of 78.68 on the MTEB code benchmark against 68.76 for EmbeddingGemma 1, with multilingual text results about the same (61.36 against 61.15), and says it leads sub-1B multimodal embedders. Those are Google's own results. Because it shares its tokenizer and audio encoder with Gemma 4, the two can run together with a smaller combined memory footprint.
Good for
- On-device and offline semantic search and retrieval-augmented generation
- Cross-modal search, such as finding a video clip from a voice memo or an image from a text query
- Local code search and indexing a codebase for coding agents
- Keeping vector storage small with truncated 256- or 128-dimension vectors
Watch out for
- Vectors are not interchangeable between embedding models, so switching models means re-indexing your data
- The 8,192-token context is short for long documents; split them into chunks
- Text works best with the recommended task prefixes; images, video and audio take none
- Benchmark results are Google's own; compare it with larger embedding models on your own data
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of EmbeddingGemma 2. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Text and code only (270M of the same weights) BF16 | 270M | – | 0.621 GB | Laptop or phone |
| Full multimodal (740M) BF16 | 740M | 1.5 GB | 1.72 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Qwen3-Embedding vs EmbeddingGemma 2: which should you use?
- Best open-source embedding models in 2026
- EmbeddingGemma 2 embeds text, images, video and audio under Apache 2.0
Frequently asked questions
- What is EmbeddingGemma 2?
- An embedding model turns text, pictures, video or sound into a list of numbers, so that things with similar meaning end up close together. That is what powers semantic search and retrieval-augmented generation. EmbeddingGemma 2 does this for all of those kinds of content in one shared space, so you can search video with a voice memo or find photos with a sentence. It is small enough to run on a phone or laptop, and the license is Apache 2.0.
- Can I use EmbeddingGemma 2 commercially?
- Yes. EmbeddingGemma 2 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is EmbeddingGemma 2 open source?
- EmbeddingGemma 2 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run EmbeddingGemma 2 on my own computer?
- Yes. The smallest variant, Text and code only (270M of the same weights), needs about 0.621 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was EmbeddingGemma 2 released?
- EmbeddingGemma 2 was released by Google DeepMind on October 6, 2026.
- How long a context window does EmbeddingGemma 2 have?
- EmbeddingGemma 2 supports a context window of up to 8K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Google blog: EmbeddingGemma 2 (6 Oct 2026) blog.google
- Google Developers Blog: EmbeddingGemma 2, the developer guide developers.googleblog.com
- EmbeddingGemma 2 model card (Hugging Face) huggingface.co
- EmbeddingGemma documentation (Google AI for Developers) ai.google.dev
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-Embedding Alibaba Qwen | 600M – 7.57B | 32K | Apache 2.0 (Permissive) | 1.38 GB Laptop or phone | |
| Qwen3-VL-Embedding Alibaba Qwen | 2.13B – 8.14B | 32K | Apache 2.0 (Permissive) | 4.9 GB Laptop or phone | |
| Granite Embedding R2 IBM | 48M – 312M | 32K | Apache 2.0 (Permissive) | 0.115 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone |