Skip to content
NewsModel release

Google ships EmbeddingGemma 2, an open 740M-parameter multimodal embedder

· by Pondero Newsdesk

The short version

EmbeddingGemma 2 maps text, code, images, audio, and video into one embedding space, runs offline on a phone, and gained an 8K-token context window under Apache 2.0.

Google ships EmbeddingGemma 2, an open 740M-parameter multimodal embedder

Google folded five separate embedding jobs, text, code, images, audio, and video, into one model small enough to run offline on a phone. EmbeddingGemma 2, released October 6, replaces a text-only predecessor that Google said passed 20 million downloads since last year.

What Google released

EmbeddingGemma 2 has 740 million parameters, is built on the Gemma 4 architecture, and ships under an Apache 2.0 license, per Google's announcement. The design is modular: text-only workloads need as little as 270 million parameters, with optional 170-million-parameter vision and 300-million-parameter audio encoders added for full multimodal support. Matryoshka Representation Learning lets developers truncate output vectors from 768 dimensions down to 512, 256, or 128, cutting local vector-database storage by up to 6x, Google said. The context window grew to 8,000 tokens, four times the prior version, enough to process roughly 5.5 minutes of audio or 29 images in one pass. On code retrieval, Google reported a 9.92-point jump on MTEB Code, from 68.76 to 78.68, and said the model leads other sub-1B multimodal embedders on its own benchmark suite. Weights are live on Hugging Face and Kaggle, with deployment paths through MediaPipe, LiteRT, transformers.js, llama.cpp, Ollama, vLLM, and LMStudio.

One model replaces a stack of specialist embedders

Teams building offline search, finding a video clip from a voice memo, or indexing a document library that can't leave a device, previously needed a separate embedding model per data type, each with its own memory budget. A quantized EmbeddingGemma 2 needs roughly 191MB of active RAM for text-only use and about 567MB for the full multimodal model on a Pixel 11 Pro, per Google, small enough to fit alongside other on-device AI workloads. That shrinks the engineering cost of shipping local, privacy-preserving multimodal RAG in a mobile app, and it gives developers a credible open alternative to cloud-only embedding APIs from OpenAI and Cohere for use cases where data cannot leave the device or where offline operation is a requirement, not a preference.

What to watch next

Gemini Enterprise Agent Platform Model Garden availability is "coming soon," per Google's post, with no specific date given. Independent benchmark runs against OpenAI's and Cohere's embedding models, rather than Google's own suite, would show whether the sub-1B multimodal lead holds up outside Google's test set.

Sources