Google DeepMind Launches EmbeddingGemma 2, Its First Open, Native Multimodal Embedding Model
Google DeepMind has launched EmbeddingGemma 2, its first open, native multimodal embedding model, with weights publicly available on Hugging Face and Kaggle, licensed under Apache 2.0 for commercial fine-tuning. The model is built on the Gemma 4 architecture, mapping text, code, images, videos, and audio into the same 768-dimensional vector space.
The full model has 740 million parameters and can be disassembled as needed: core text and code at 270 million (with a backbone of 130 million and embedding layer of 140 million), visual encoder at 170 million, and audio encoder at 300 million. Developers can load only text or combine to create models of 440 million, 570 million, or 740 million parameters. The context window is 8192 tokens, covering over 100 languages, and the audio side can handle approximately 5.5 minutes of recording.
After quantization on the Pixel 11 Pro, the active memory for pure text weights is about 191 MB, while the full multimodal model is approximately 567 MB. The output supports Matryoshka truncation, allowing dimensions to be reduced from 768 to 512, 256, or 128, with Google stating that local vector storage costs can be reduced to about one-sixth. It shares the text tokenizer and audio encoder with Gemma 4, resulting in lower total memory when running both ends simultaneously.
Compared to the previous generation pure text EmbeddingGemma, code retrieval MTEB increased from 68.76 to 78.68, a rise of 9.92 points, or about 14%; the multilingual text MTEB v2 is 61.36, up from 61.15 in the previous generation. The average for image MIEB lite is 64.64, video MMEB v2 hit rate is 50.67, audio MSEB retrieval is 69.54, and MAEB average is 49.39. Google claims it leads in quality among multimodal embeddings under 1 billion parameters, surpassing some dedicated models by more than twice the size.
The previous generation EmbeddingGemma has been downloaded over 20 million times, primarily for edge-side search and privacy-focused retrieval-enhanced generation. The new model can locate video clips by voice memo or retrieve hours of audio using text, with data not needing to leave the device. Deployment channels also include Ollama, llama.cpp's GGUF and LiteRT, and browser-side can utilize transformers.js or WebGPU.
Edge-side retrieval shifts purchases from cloud-based embedded APIs to open-source weights and local computing power. Application developers, terminal manufacturers, and privacy-sensitive industries are the direct users, putting pressure on cloud embedding services and larger parameter dedicated models. The event-driven nature stems from the extension of the Gemma 4 open weight system, with funding and traffic flowing towards Hugging Face weights, edge-side inference frameworks, and local vector libraries, rather than per-call cloud APIs.
Source: Public Information
ABAB AI Insight
Google's embedding strategy is to first collect in the cloud and then deploy on the edge. The EmbeddingGemma launched in September 2025 is a 308 million pure text model, focusing on leading MTEB under 500 million, with over 20 million downloads; earlier general embedding capabilities were centered around paid APIs like Gemini Embedding, with outputs reaching 3072 dimensions. EmbeddingGemma 2 compresses the text core to 270 million while adding visual and audio encoders, effectively compensating for the previous generation's inability to perform cross-modal retrieval with open weights. The Gemma series has also faced licensing tightening controversies, with developers questioning the gap between 'open' and commercial terms; this shift to Apache 2.0 addresses that historical friction.
The capital path is not through acquisitions but by separating the retrieval layer from cloud billing. Google's own Gemini Embedding 2 remains a paid API, with no public edge-side memory data; EmbeddingGemma 2 releases weights, GGUF, and LiteRT together, clearly sharing tokenizers and audio encoders with Gemma 4, facilitating a single edge-side pipeline for 'small model retrieval, large model inference.' The motivation is to maintain the default search entry for Android, Pixel, and Chrome: if developers build a vector library on their phones, subsequent generation calls are more likely to stay within Gemma or Gemini rather than flow to external embedding services.
Comparable objects should be viewed separately. Alibaba's Qwen3-VL-Embedding-2B has about 2 billion parameters, reporting MMEB-v2 at about 73.2, without audio; research-oriented LCO-Embedding-Omni-3B has a larger backbone, covering audio and video. A historical comparison of similar size is the previous generation EmbeddingGemma catching up to nearly twice the parameter text models. The industry position has moved past 'can it embed' to the control phase of 'must retrieval go to the cloud': under 1 billion parameters, truncatable dimensions, and offline capabilities are shifting the entry of RAG from API vendors to device-side components.
This is a transfer of pricing power, not merely a technical replacement. When embedding is billed by token, long audio, code libraries, and video frames are all usage for cloud vendors; once the 768 dimensions can be truncated to 128 dimensions, storage reduced to about one-sixth, and the full multimodal model running at about 567 MB can operate on a phone, the marginal cost of retrieval returns to the device and electricity. The mechanism is that small model quality crosses the usability line, while open-source licenses allow manufacturers to pre-install. Cloud APIs still retain closed-source quality and ultra-long context, but the default privacy retrieval and code library indexing are beginning to shift pricing power from call frequency to who controls the terminal.
ABAB News · Cognitive Laws
Retrieval is free, then generation starts charging.
Small models win not by parameters, but by being closer to the data.
Open weights are the entry point, closed-source APIs are the cash register.