Choose Rich Live
Tech

Google releases EmbeddingGemma 2 multimodal embedding model for local devices

Google has launched EmbeddingGemma 2, an open-weight 740-million-parameter multimodal embedding model that processes text, code, audio, images, and video on-device under an Apache 2.0 license.

Google and Google DeepMind have launched EmbeddingGemma 2, an open-weight 740-million-parameter embedding model distributed under an Apache 2.0 license. The model is designed to operate locally on user hardware such as phones, laptops, and glasses, reducing cloud reliance for personal AI applications. It maps text, code, images, audio, and video into a single unified embedding space, allowing developers to execute multimodal search directly without relying on separate transcription or captioning workflows.[1][2][3][4][5][6]

The release includes modular encoders and an 8K context window, as well as a compact text-only mode that reduces the model to 270 million parameters. Utilizing Matryoshka Representation Learning, EmbeddingGemma 2 can achieve up to a sixfold reduction in storage requirements. When tested on a Pixel 11 Pro, memory consumption registered at approximately 191MB of RAM for text-only weights and roughly 567MB for the complete multimodal model.[5][6]

Targeting local workloads such as retrieval-augmented generation (RAG), media search, video retrieval, and semantic indexing, the model supports runtime environments including MLX, Ollama, LM Studio, llama.cpp, vLLM, and transformers.js.[5]

Key facts

  • Google DeepMind released EmbeddingGemma 2 as an open-weight, 740-million-parameter model under an Apache 2.0 license.
  • The model maps text, code, images, audio, and video into a shared embedding space without requiring captioning or transcription steps.
  • EmbeddingGemma 2 features an 8K context window, modular encoders, and a text-only configuration scaled to 270 million parameters.
  • Matryoshka Representation Learning enables up to a 6x reduction in storage space.
  • Hardware memory requirements on a Pixel 11 Pro are approximately 191MB of RAM for text-only weights and about 567MB for the full multimodal model.
  • Supported tools and runtimes include MLX, Ollama, LM Studio, llama.cpp, vLLM, and transformers.js.

Sources · 6 sources

  1. CB

    Crypto BriefingArticle ·

    Google releases EmbeddingGemma 2, a 740 million parameter open-weight model under Apache 2.0 EmbeddingGemma 2's release under Apache 2.0 democratizes AI development, enabling privacy-focused, on-device applications and reducing server reliance. The post Google releases EmbeddingGemma 2, a 740 million parameter open-weight model under Apache 2.0 appeared first on Crypto Briefing .

    Open source
  2. TN

    The New Stack@thenewstackPost on X ·

    Google's EmbeddingGemma 2 brings text, code, image, video and audio search to one on-device model, so developers can skip captioning and transcription. https://t.co/oRko43GD0J

    Open source
  3. TE

    Techmeme@TechmemePost on X ·

    Google DeepMind launches EmbeddingGemma 2, a 740M-parameter model to map code, images, video, and audio in a shared embedding space, under an Apache 2.0 license (Google) (Visit Techmeme dot com for the link and full context!)

    Open source
  4. SB

    Shay Boloor@StockSavvyShayPost on X ·

    $GOOGL just released EmbeddingGemma 2 which is a 740M parameter open model that handles code, images, video & audio entirely on device. That pushes more AI from cloud onto phones, laptops & glasses as personal agents become more private & always on. https://t.co/pJ4HcNSURa

    Open source
  5. AS

    AshutoshShrivastava@ai_for_successPost on X ·

    🚨 Massive from Google... They just dropped EmbeddingGemma 2, a 740M parameter multimodal embedding model that runs locally. > Maps text, images, audio, video and code into one unified embedding space > 740M parameters, with text-only mode as small as 270M > Apache 2.0 licensed > 8K context window > Up to 6x storage reduction with Matryoshka Representation Learning > ~191MB RAM for text-only weights and ~567MB for the full multimodal model on a Pixel 11 Pro > Supports on-device semantic search, RAG, media search and video retrieval > Works with MLX, Ollama, LM Studio, llama.cpp, vLLM, transformers.js and more This is a pretty big deal for local multimodal AI. https://t.co/IFupRr9rFo

    Open source
  6. �A

    🚨 AI News | TestingCatalog@testingcatalogPost on X ·

    Google released EmbeddingGemma 2 open weight model under Apache 2 license! > “Our lightweight, multimodal embedding model maps text, code, images, video, and audio into a single, unified embedding space.“ > 740M parameter form factor with modular encoders and 8K context window. Embedded testing time 👀 https://t.co/Z5qrbp1Xk1

    Open source