EmbeddingGemma 2: an open, lightweight multimodal embedding model

What it claims

EmbeddingGemma 2 is a 740M-parameter open embedding model under Apache 2.0, built on Gemma 4. It maps text, code, images, video and audio into one shared vector space. The page gives one hard benchmark delta against its predecessor: MTEB Code rises from 68.76 (EmbeddingGemma 1) to 78.68, a 9.92-point gain. Google says multilingual text quality is unchanged from v1. Google positions it for local codebase indexing, semantic code search and coding-agent retrieval.

How it works

The 740M total is three separable parts: a 270M text backbone, an optional 170M vision encoder and an optional 300M audio encoder. You load only the encoders your workload needs.

flowchart LR
  T[Text/code input] --> B[Text backbone 270M]
  I[Image/video input] -.optional.-> V[Vision encoder 170M]
  A[Audio input] -.optional.-> AU[Audio encoder 300M]
  V --> B
  AU --> B
  B --> E[Shared embedding, 768 to 128 dims]

The 8K context (4× v1) translates to concrete per-input capacity:

• up to 5.5 minutes of audio,
• 29 images,
• 58 video frames,
• or interleaved mixes of these.

The model shares Gemma 4's text tokenizer and audio encoder. Running it alongside Gemma 4 in an on-device RAG pipeline therefore costs less total memory than two independent models would. MRL truncation points are 768, 512, 256 and 128 dimensions, up to 6× storage reduction.

How to use it

• Weights: Hugging Face and Kaggle. On-device-optimised builds are under LiteRT Community on Hugging Face. Gemini Enterprise Agent Platform Model Garden is "coming soon".
• On-device: use MediaPipe (Google AI Edge) for turnkey embedding, retrieval and decision tasks. Use LiteRT for custom integration. Browser: transformers.js or WebGPU.
• Serving: transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LMStudio.
• Vector store: Qdrant is the named partner.
• Fine-tuning: guidance from Unsloth.

The page gives no install or run commands. Those are in the linked developer guide and model card.

Limits

• The only number with a baseline is MTEB Code. MAEB, vision, video and document claims ("leading", "outperforms some specialist models more than twice its size") have no scores or named comparators on this page. Check the model card before relying on them.
• The RAM figures (~191MB text-only, ~567MB full multimodal) are measured on a Pixel 11 Pro with quantization. The quantization scheme and the latency are not stated.
• No storage or quality cost is given for each MRL truncation level.
• This is a launch post. It does not discuss failure modes or untested settings.