EmbeddingGemma 2 Released: A New Benchmark for On-Device Multimodal Embedding Models

Google DeepMind has introduced EmbeddingGemma 2, expanding embedding model capabilities from pure text to code, images, video, and audio — all packed into a model with just 740 million parameters that can run directly on a smartphone. This marks a significant leap forward for on-device AI retrieval.

From Text to Multimodal

The original EmbeddingGemma, released last year, was positioned as a lightweight, high-quality text embedding solution — helping developers organize, search, and connect information directly on consumer hardware. Its market reception exceeded expectations: over 20 million downloads, widely adopted in intelligent on-device search tools and privacy-conscious retrieval-augmented generation (RAG) pipelines.

The core upgrade in EmbeddingGemma 2 is native multimodality. Built on the Gemma 4 architecture and released under the commercially friendly Apache 2.0 license, it maps text, code, images, audio, and video into a single unified embedding space. This means developers can use a voice memo to find a specific video clip, or use a text query to locate moments within hours of recordings — all handled natively by a single model, with no need to chain together multiple specialized models.

What is a "unified embedding space"? Embeddings are the process of mapping different content into high-dimensional numerical vectors, where the distance between vectors reflects semantic similarity. Single-modality models can only compare content of the same type (text vs. text), while multimodal models project different modalities into the same coordinate system — enabling a text description and a video frame to have their similarity directly computed. Retrieval-Augmented Generation (RAG) is a common AI application pattern: before generating a response, an embedding model retrieves relevant passages from a knowledge base, which are then fed as context to the generative model — reducing hallucinations and improving accuracy and timeliness. On-device RAG is especially valuable because the entire retrieval process requires no internet connection; data never leaves the device, making it ideal for privacy-sensitive use cases.

Modular Design and Storage Efficiency

EmbeddingGemma 2 carries forward the technical DNA of the Gemini Embedding family, while incorporating extensive engineering optimizations for on-device scenarios.

Modular Architecture with On-Demand Loading

The model uses a modular design: text-only tasks can run with as few as 270 million parameters, while the vision encoder (170M parameters) and audio encoder (300M parameters) are optional components that can be enabled as needed. This lets developers trim resource usage based on actual requirements, avoiding the memory cost of modalities they don't use.

Storage Compression via Matryoshka Representation Learning

Leveraging Matryoshka Representation Learning (MRL), developers can dynamically truncate output vectors from 768 dimensions down to 512, 256, or even 128 dimensions. This enables up to 6× storage compression for local vector databases and memory usage — a significant benefit for resource-constrained on-device hardware.

After quantization, the model's real-world performance on the Google Pixel 11 Pro is impressive: the text-only weights require roughly 191MB of active memory, while the full multimodal model needs only about 567MB. This footprint makes it genuinely viable for persistent on-device deployment.

What is Matryoshka Representation Learning (MRL)? MRL is a vector training technique inspired by Russian nesting dolls. Traditional embedding models output fixed-dimension vectors that must be used at full size. MRL, by contrast, simultaneously optimizes sub-vectors at multiple truncated dimensions during training — so the same model's output can be trimmed on demand at different precision levels. For example, the first 128 dimensions of a 768-dimensional vector already form an independently usable embedding of acceptable quality; the first 256 dimensions offer higher quality; and so on. This "nested" structure lets developers flexibly trade off storage, retrieval speed, and retrieval precision: use low-dimensional vectors for fast candidate pre-filtering, then switch to high-dimensional vectors for precise re-ranking. The storage compression ratio scales linearly with dimension reduction — going from 768 to 128 dimensions is exactly 6×, which directly corresponds to the "up to 6× storage compression" mentioned above.

Performance: Punching Above Its Weight

EmbeddingGemma 2 maintains comparable multilingual text capabilities to the original while achieving a significant jump in code understanding. On the MTEB Code benchmark, its score improved from 68.76 to 78.68 — a gain of 9.92 points. This makes it well-suited for local codebase indexing, semantic code search, and retrieval in coding agents.

Across image, video, document, and audio dimensions, EmbeddingGemma 2 sets a new standard for "quality per parameter" among sub-billion-parameter models, even outperforming some specialized models twice its size. On benchmarks including MTEB Code and MAEB (Massive Audio Embedding Benchmark), it achieves leading scores among multimodal embedding models of its size class.

Another notable upgrade is the context window. EmbeddingGemma 2 supports an 8K token context — four times that of the original — enabling it to natively process up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations of these on local hardware.

About the benchmarks: MTEB (Massive Text Embedding Benchmark) is the authoritative comprehensive benchmark for evaluating text embedding models, covering retrieval, classification, clustering, semantic similarity, and more — widely used in the industry for cross-model comparisons. MTEB Code is its code-focused subset, testing a model's ability to understand and retrieve functions, comments, and code semantics. MAEB (Massive Audio Embedding Benchmark) is the audio equivalent, evaluating semantic representation across speech, sound effects, music, and other audio types. Score improvements on these benchmarks directly demonstrate that EmbeddingGemma 2 outperforms same-size competitors on specialized domain retrieval tasks — not merely on general text capability.

Real-World On-Device Semantic Retrieval Use Cases

EmbeddingGemma 2 brings powerful retrieval capabilities directly to edge hardware. Generating embeddings locally ensures data privacy, reduces pipeline latency, and enables cross-modal search to run entirely offline.

When paired with a generative model like Gemma 4, it can power on-device RAG pipelines that understand complex multimodal data. Since EmbeddingGemma 2 is itself built on Gemma 4 and shares its text tokenizer and audio encoder, developers can run both models together in a unified pipeline with a lower combined memory footprint.

Google has also highlighted several ready-to-experience application scenarios:

• Instant Media Search in Google AI Edge Gallery: search a media library by semantic similarity using text or images
• Video Moments Finder: locate specific moments in a video using text or audio queries
• Google AI Edge Foresight: combine EmbeddingGemma 2's local file retrieval with Gemma 4's contextual reasoning
• MediaPipe Decision Task API: build real-time classification, routing, and prediction engines that leverage multimodal context

Getting Started

EmbeddingGemma 2 has strong ecosystem support. Model weights are available on Hugging Face and Kaggle, with support coming soon to the Gemini Enterprise Agent Platform Model Garden. On-device optimized versions are available via LiteRT Community on Hugging Face.

For deployment, developers can use Google AI Edge MediaPipe for out-of-the-box embedding, retrieval, and decision tasks; LiteRT for custom integrations; or transformers.js with WebGPU to run it in the browser. The model is compatible with major toolchains including transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio. Vector storage can be paired with Qdrant, and fine-tuning guides are available via Unsloth.

Conclusion

EmbeddingGemma 2 represents an important milestone in on-device multimodal retrieval. With fewer than one billion parameters, it delivers unified embedding across text, code, images, audio, and video — maintaining leading performance while preserving privacy and enabling fully offline operation. For developers looking to build search, RAG, and decision systems on consumer-grade devices, it's an open-source option that balances both practicality and accessibility.