EmbeddingGemma 2 launched on 6 October 2026 with an 8,192-token context window, four times its predecessor's, and needs about 191MB of RAM for text-only use on a Pixel 11 Pro. It suits teams building on-device search and RAG over mixed media, and replaces EmbeddingGemma 1 wherever images, audio, or video need indexing.
What changed
- · Release EmbeddingGemma 2 released as an open multimodal embedding model source
EmbeddingGemma 2 is Google DeepMind's open embedding model that places text, code, images, video, and audio in one 768-dimension vector space, scoring 78.68 on MTEB Code, up from 68.76 for its predecessor. At 740 million parameters it is built to run on phones and laptops under an Apache 2.0 license.
Provider: Google DeepMind · Family: EmbeddingGemma
More about Google DeepMind on HokAI
Context window: 8,192 tokens
Input modalities: text, code, image, audio, video · Output: embeddings
About EmbeddingGemma 2
EmbeddingGemma 2 is an open multimodal embedding model from Google DeepMind, released on 6 October 2026. It turns text (including code), images, video, and audio, alone or mixed in one input, into a single 768-dimension vector, so a spoken query can retrieve a video clip and a text query can retrieve an image. It carries 740 million parameters in total and is licensed under Apache 2.0. It is the second model in the EmbeddingGemma line: the first was text-only and passed 20 million downloads, according to Google's launch post.
The architecture is modular. A 270M-parameter text backbone (130M transformer plus a 140M embedder) is always loaded, and the 170M vision encoder and 300M audio encoder can be switched off, which gives effective sizes of 270M for text only, 440M for text and image, 570M with audio, and 740M for everything. The backbone has 24 layers, a model dimension of 512, a hidden dimension of 2,048, a sliding window of 1,024 tokens with a 5:1 local-to-global attention ratio, a 262,144-token vocabulary, and mean pooling followed by a 512 to 768 projection. It builds on the Gemma 4 architecture and shares the Gemma 4 text tokenizer and audio encoder, so running it beside a Gemma 4 generator costs less combined memory than two unrelated models. Larger siblings such as Gemma 4 31B generate text; this model only produces vectors.
Google reports its results at the full 768 dimensions. On the multilingual MTEB v2 text benchmark it scores 61.36 against 61.15 for EmbeddingGemma 1, which is a tie in practice. The jump is in code: MTEB Code (NDCG@10) rises from 68.76 to 78.68, a 9.92-point gain. For the new modalities there is no predecessor to compare with: 65.67 on MIEB lite for images, 57.28 on MMEB v2 Image, 67.84 on MMEB v2 VisDoc for document pages, 50.67 on MMEB v2 Video, 69.54 on MSEB Retrieval for audio, and 49.87 on MAEB. All of these are vendor-reported on the model card; HokAI has not re-run them.
The context window is 8,192 tokens, four times the 2,048 of EmbeddingGemma 1, and every modality draws from that single budget. At the default settings an image costs 280 tokens (about 29 images per input), a video frame costs 140 tokens (about 58 frames at one frame per second), and audio costs 25 tokens per second (about 327 seconds). Lowering the vision budget raises the image or frame count to about 114. Output vectors can be cut to 512, 256, or 128 dimensions through Matryoshka training. At 256 dimensions MTEB multilingual drops from 61.36 to 60.41 and MTEB Code from 78.68 to 76.18, while storage falls to a third. At 128 dimensions the multimodal scores fall sharply (MMEB v2 overall goes from 59.01 to 45.65), so Google advises 128d only for text workloads.
The model targets on-device work: local search over a photo or video library, retrieval for RAG on a phone or laptop, code search over a repository, and clustering or classification. Google states that with quantization on a Pixel 11 Pro it needs about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model. Weights are on Hugging Face and Kaggle, a Gemini Enterprise Agent Platform Model Garden listing is marked as coming soon, and Google names support in transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LM Studio, Qdrant, MediaPipe, and LiteRT. Google does not sell hosted inference for it, so the working cost is your own hardware.
Three practical limits matter. Run it in bfloat16 or float32, never float16: the card says activations exceed float16's range and the output becomes NaN or quietly degraded rather than raising an error. Prepend the task prefix (for example "task: search result | query: ...") to text inputs, because omitting it lowers quality; images, video, and audio take no prefix. And re-normalize after truncating dimensions, since a sliced unit vector no longer has unit length. It is a pre-trained embedder with no post-training alignment, and the card says safeguards such as retrieval filtering are the deployer's job. Pre-training data has a January 2025 cutoff and covers 140+ languages, with 100+ claimed as supported at unequal quality.
Who should pick it: teams that want one small open model to index text, code, images, video, and audio on a phone or laptop under a permissive license, especially those already using Gemma 4 12B for generation. Who should look elsewhere: teams that need documents longer than 8,192 tokens in a single vector, a vendor-hosted API with an SLA, or independently verified rankings, since the numbers above come from Google. Google's own larger Gemini Embedding models share its technology and suit server-side use, and other Google DeepMind models such as Gemini Nano cover the on-device generation side. For comparison, other small multimodal models in the directory include LFM2.5 VL 3B, Gemini Omni, Gemini 3.5 Flash Lite, and DiffusionGemma 26B A4B; they generate rather than embed, so they pair with an embedder instead of replacing one. Not sure which model fits your stack? Smart Match asks a few questions and recommends one.
Pricing
EmbeddingGemma 2 has no per-token fee: the Apache-licensed download costs nothing and Google does not sell hosted inference for it. A Model Garden listing is marked coming soon on the launch post, so no managed price existed on 28 October 2026.
| Tier | Rate |
|---|---|
| Self-hosted (open weights) | Free under Apache 2.0. Runs on CPU, laptop GPU, or phone; you pay only for your own hardware. |
Key Features
- Four modalities, one space: Text, code, images, video, and audio map into a single 768-dimension vector, and one input can interleave them using placeholder tokens.
- Modular encoders: The 270M text core runs alone; add the 170M vision or 300M audio encoder only when needed.
- Matryoshka dimensions: Vectors truncate to 512, 256, or 128 dimensions for up to 6x less storage.
- 8K shared context: An 8,192-token window covers about 29 images, 58 video frames, or 327 seconds of audio at default settings.
- Task-steered text embeddings: Short prefixes tune text vectors for search, question answering, classification, clustering, or similarity.
Pros
- MTEB Code score of 78.68, a 9.92-point gain over EmbeddingGemma 1 (vendor-reported).
- Permissively licensed weights with no per-token fee and no Google-operated API receiving your data.
- Quantized footprint of about 191MB (text) to 567MB (multimodal) on a Pixel 11 Pro, per Google.
Cons
- Multilingual text quality is flat against version 1 (61.36 vs 61.15 on MTEB v2).
- Float16 is unsafe, so older GPUs without bfloat16 need float32 and more memory.
- No independently verified benchmark runs exist yet, and no hosted API with an SLA.
Frequently Asked Questions
What does EmbeddingGemma 2 cost to run?
The weights are free under Apache 2.0, and there is no per-token fee because Google does not sell hosted inference for the model. Google's launch post lists a Model Garden option as coming soon, so on 28 October 2026 the only cost is your own phone, laptop, or server. Quantized text-only use needs about 191MB of RAM on a Pixel 11 Pro, per Google.
Is EmbeddingGemma 2 any better than the first EmbeddingGemma?
For code, yes: MTEB Code rose by 9.92 points. For multilingual text it is a tie, 61.36 against 61.15 on MTEB v2, so a text-only team gains little beyond the 8,192-token window, four times larger than before. The real change is that images, video, and audio can now be embedded into the same space. All figures are Google's own.
Can EmbeddingGemma 2 go into a commercial product?
Yes. It is released under Apache 2.0, which permits commercial use, and the weights are on Hugging Face and Kaggle. Deployments must still follow the Gemma Prohibited Use Policy, and Google leaves application-level safeguards such as retrieval filtering to the deployer.
Does EmbeddingGemma 2 send my data to Google?
Not when you run the weights yourself: Google operates no inference API for it, so inputs stay on your own device or server. The model card describes pre-training data (cutoff January 2025) but no user-data training, because it never sees user data in self-hosted use. If a managed Model Garden listing appears, check that service's own terms.
When should a team skip EmbeddingGemma 2?
Skip it if single documents run past 8,192 tokens, if you need a hosted API with an SLA, or if your hardware lacks bfloat16 and cannot spare float32 memory, since float16 produces NaN output. Teams wanting a server-side option from the same lab can look at Google's Gemini Embedding models. For phones and laptops indexing mixed media, it is the better fit.