embeddinggemma-2: 4GB+ VRAM, File List

embeddinggemma-2: 4GB+ VRAM, File List

At a Glance

Item Value
Repository google/embeddinggemma-2
Publisher guide Google: models and licenses
Published 2026-09-14
License apache-2.0
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Google DeepMind has released “EmbeddingGemma 2", an open multimodal embedding model capable of mapping four modalities—text, image, audio, and video—into a single vector space. Despite its lightweight design with 740 million parameters, the model enables advanced on-device semantic search, RAG (Retrieval-Augmented Generation), and information organization.

The previous model, EmbeddingGemma, recorded over 20 million downloads and has been utilized for privacy-focused on-device search. This new model natively integrates not only text but also code, images, video, and audio, enabling cross-modal processing such as “finding a specific video clip from an audio memo" in a local environment with a single model.

Specifications

The main specifications listed in the documentation are as follows:

  • Parameter Count & Architecture: A total of 740 million parameters. Based on the Gemma 4 architecture, it combines a 270 million parameter text backbone (130 million transformer and 140 million embedding layers) with encoders for images (170 million) and audio (300 million).
  • Modular Design: Encoders can be loaded individually depending on the use case. Flexible operation is possible, such as 270 million for text only, 440 million for text and images, or the full set of 740 million.
  • Output Specifications: Natively outputs 768-dimensional vectors. It adopts Matryoshka Representation Learning (MRL), allowing dynamic truncation to 512, 256, and 128 dimensions while minimizing the impact on quality.
  • Context Length: 8,192 tokens. This is expanded to 4 times the previous model, allowing it to process several minutes of audio and numerous images at once.
  • Supported Media Specifications:
  • Text: Supports over 100 languages. The use of task-specific instruction prefixes (SearchQuery, Classification, etc.) is recommended. – Image: Consumes 280 tokens per image, supporting up to approximately 29 images. – Video: Default sampling at 1fps, consuming 140 tokens per frame (up to approximately 58 frames). – Audio: Consumes 25 tokens per second, supporting up to approximately 327 seconds. 16kHz mono format is recommended.
  • License: Apache 2.0 (a permissive license allowing commercial use)

Performance and Quality

EmbeddingGemma 2 achieves class-leading scores across multiple benchmarks as a multimodal embedding model with fewer than 1 billion parameters. Code-related performance in particular has improved significantly, achieving an approximately 14% improvement on the MTEB Code benchmark from 68.76 in the previous model to 78.68.

The main evaluation results published by the authors are as follows:

→ Scroll horizontally to see all columns

Modality Benchmark Metric EmbeddingGemma 2 EmbeddingGemma 1
Text Massive Text Embedding Benchmark (MTEB, multilingual, v2) Mean(Task), Multiple 61.36 61.15
Massive Text Embedding Benchmark (MTEB, code, v1) Mean(Task), NDCG@10 78.68 68.76
Image Massive Image Embedding Benchmark (MIEB, lite) Mean(TaskType), Multiple 64.64 –
Massive Multimodal Embedding Benchmark (MMEB v2 – Image) Mean(Task), Hit@1 57.28 –
Massive Multimodal Embedding Benchmark (MMEB v2 – VisDoc) Mean(Task), NDCG@5 67.84 –
Video Massive Multimodal Embedding Benchmark (MMEB v2 – Video) Mean(Task), Hit@1 50.67 –
Audio Massive Sound Embedding Benchmark (MSEB, Retrieval) Mean(Task), MRR@10 69.54 –
Massive Audio Embedding Benchmark (MAEB) Hugging Face Mean(Task), Multiple 49.39 –

In terms of text performance, while maintaining high multilingual capability equivalent to the previous model, it demonstrates a noticeable advantage in code retrieval. Furthermore, in each modality of image, video, and audio, it is reported to achieve a quality that surpasses specialized models more than twice its size.

In evaluations regarding vector dimensionality reduction (MRL), the preservation rate of accuracy when reducing dimensions is shown below:

→ Scroll horizontally to see all columns

Output Dimension Compression Ratio MTEB (multilingual) MTEB (code) MIEB (Image) MMEB (Overall) MSEB (Audio)
768d (Full) 1:1 61.36 78.68 64.64 59.01 69.54
512d 1:1.5 61.17 77.24 64.32 58.38 69.18
256d 1:3 60.41 76.18 63.13 56.24 66.76
128d 1:6 57.89 71.41 59.06 45.65 56.71

Up to 256 dimensions, the drop in accuracy is extremely small, enabling practical searches while reducing storage capacity by a third. On the other hand, reducing down to 128 dimensions results in a significant performance drop, particularly in multimodal tasks, so verification according to the use case is recommended.

Strengths and Use Cases

EmbeddingGemma 2 demonstrates its true value in multimodal semantic search and RAG (Retrieval-Augmented Generation) in on-device environments. Because it can integrate text, code, images, video, and audio into a single vector space, it enables complex retrievals that were difficult with previous text-centric models.

As specific intended use cases, the authors list the following:
– Cross-modal retrieval: It is possible to search within audio archives using text queries, or to identify related video scenes using images as input. For example, usage such as “finding a specific video clip from the content of an audio memo" is envisioned.
– Advanced code search and indexing: Due to the high performance demonstrated in the MTEB Code benchmark, it is suitable for semantic search of local codebases and retrieval for coding agents.
– Privacy-preserving on-device RAG: Because all processing can be completed offline, it is ideal for enterprise knowledge bases dealing with confidential data or information organization on personal devices.
– Classification and clustering: It can be used for tasks that classify inputs based on preset labels, such as sentiment analysis, content moderation, audio event detection, and image categorization. It is also possible to group document sets based on semantic similarity.
– Fact verification: It can help build automated fact-checking systems or source attribution pipelines by searching for evidentiary documents against claims.

Additionally, this model supports “interleaved" input. By inserting placeholders such as <|image|>, <|video|>, and <|audio|> into text, information containing multiple media types—such as “Waterproof running shoes. <|image|> Breathable mesh material. <|video|>"—can be represented as a single 768-dimensional vector. This facilitates product catalog searches and the understanding of complex multimedia documents.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 744M parameters

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) BF16 1.4GB 1.7GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Distributed Files

Weight files published in google/embeddinggemma-2, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
model.safetensors 1.49GB

How to Get It

Model weights are available from the Hugging Face repository google/embeddinggemma-2 and Kaggle. Models optimized for on-device use are also provided on the LiteRT Community Hugging Face page. Distribution formats support safetensors.

This model supports a wide development ecosystem and is immediately available in the following tools and libraries:
– Python libraries: Can be easily integrated using sentence-transformers and transformers.
– Inference engines and tools: Supported by many tools familiar to local LLM users, such as llama.cpp, Ollama, LMStudio, vLLM, MLX, and SGLang.
– Mobile and web deployment: Cross-platform app development using Google AI Edge MediaPipe or inference in the browser using transformers.js and WebGPU is possible.
– Vector databases: Generated vectors can be stored in databases like Qdrant for efficient searching.

Important points to note during implementation include numerical precision and normalization handling:
– Numerical precision: Inference must be performed in bfloat16 or float32. float16 cannot fully cover the activation range of the model, resulting in degraded quality without throwing errors or returning NaNs, so it should not be used. bfloat16 is recommended if hardware supports it.
– Task prefixes: When inputting text, attaching task-appropriate prefixes optimizes embedding quality. For asymmetric tasks like search, use SearchQuery on the query side and Document on the document side. For symmetric tasks like classification and similarity measurement, apply the same prefix such as Classification or SentenceSimilarity to all inputs being compared.
– Document formatting: Documents with a title are recommended to be inputted in the format title: {title} | text: {content}, and without a title as title: none.
– Post-processing after dimensionality reduction: After truncating vectors using Matryoshka Representation Learning (MRL) (e.g., from 768d to 256d), always re-run L2 normalization before calculating cosine similarity. These processes can be automated by specifying truncate_dim and normalize_embeddings=True when calling model.encode().

Related Articles

What to Read Next

Sources