EmbeddingGemma 2: on-device multimodal embeddings under Apache 2.0

Google DeepMind released EmbeddingGemma 2 on October 6, an open-weight model that maps text, images, video frames, and audio into one shared 768-dimensional vector space. According to Google, it has 740 million total parameters, ships under Apache 2.0, and needs about 191 MB of active RAM for text-only work with quantization on a Pixel 11 Pro. Weights are on Hugging Face and Kaggle, with integrations from sentence-transformers and llama.cpp to LiteRT and MediaPipe.

Composite screenshot: Google blog headline and hero (top) and the Hugging Face model card vector-truncation table (bottom). Source: blog.google and huggingface.co · Capture date: 2026-10-07 · Credit: Google DeepMind · Apache 2.0 weights; screenshots used editorially.

What happened

Google’s announcement describes EmbeddingGemma 2 as a model that “natively maps combinations of text, images, audio, and video into a unified embedding space,” replacing pipelines that chain separate captioning, speech-to-text, and text-embedding models (Google announcement, 2026-10-06). It is built on the Gemma 4 architecture.

The model is modular (model card, Hugging Face):

Google reports two quantized footprints on a Pixel 11 Pro: about 191 MB active RAM for text-only and about 567 MB for the full multimodal model — with quantization, not full precision (Google announcement).

An 8,192-token context window — four times EmbeddingGemma 1 — holds about 5.5 minutes of audio, 29 images, or 58 video frames. Matryoshka Representation Learning (MRL) truncates the 768-dimensional output to 512, 256, or 128 dimensions; Google cites up to 6x storage reduction, and the AI Edge blog up to 8x (announcement, AI Edge).

Google’s headline quality number is code retrieval: MTEB Code (v1) rises from 68.76 for EmbeddingGemma 1 to 78.68, a 9.92-point gain. The model card reports the full set (model card):

Why it matters

The practical change is index consolidation: one index holds and cross-queries text, images, and audio, so teams stop maintaining separate text and vision retrieval stacks for local search (model card). The same model can also index a codebase for semantic code search and coding-agent retrieval (announcement).

Practical implications

• On-device app developers can use one model for privacy-first RAG, search, clustering, and classification without cloud APIs. A text-only app loads just the 270M backbone.
• Agent platform engineers can build offline retrieval. Google says: “When paired with generative models such as Gemma 4, EmbeddingGemma 2 enables on-device RAG pipelines.” It shares Gemma 4’s text tokenizer and audio encoder.
• Builder and tool authors can integrate through transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LMStudio, transformers.js/WebGPU, and Qdrant, or use LiteRT and MediaPipe Tasks for turnkey on-device embeddings.

For local vector stores, MRL shows where truncation is cheap and where it is not (Hugging Face):

Risks and caveats

• Vendor-reported numbers. Parameters, dimensions, latency, and benchmarks come from Google, and the published benchmarks use the full-precision checkpoint, not the quantized phone runtime.
• “Best-in-class” is positioning. Google’s “leading scores among sub-1B multimodal embedders for its size” claim covers MTEB Code and MAEB, with no published comparator table and only Google’s own full-precision runs.
• MRL cost. Quality is near-lossless to 256 dimensions, but 128 dimensions “degrades multimodal quality substantially” and suits text-only workloads.
• Precision trap. Run inference in bfloat16 or float32; the card warns that float16 returns NaN or silently degraded embeddings.
• No safety tuning. The model card states it is a pre-trained embedding model without post-training alignment or output-level moderation.

What to watch

• Whether ML Kit for Android (“in the coming weeks”) and Model Garden (“coming soon”) get dated general availability.
• Whether independent evaluations reproduce the 9.92-point MTEB Code gain, given Google’s runs are self-reported.
• How quickly the listed runtimes and Qdrant ship tested EmbeddingGemma 2 support, and whether the modular footprint holds in real memory-constrained deployments.

Sources