EmbeddingGemma 2 is Google DeepMind's open, lightweight multimodal embedding model. It maps text, code, images, video and audio into one shared space and is small enough to run on a phone.
Problem: Searching by meaning usually needs a pipeline of separate models: a captioning model for images, speech-to-text for audio, a text embedder on top. That's slow, memory-hungry, and often sends your data to the cloud. For private, on-device apps, that's a hard trade-off.
Solution: One 740M-parameter model, built on the Gemma 4 architecture, that embeds all five types of content in the same space. A text query can find a photo, and a voice memo can find a video clip. Everything runs locally, so your data never has to leave the device.
What makes it different: It's modular. You can load only the encoders you need, as little as 270M parameters for text and code, plus optional vision (170M) and audio (300M) encoders. With quantization on a Pixel 11 Pro, it needs about 191MB of RAM for text only and about 567MB for the full multimodal model. It shares its tokenizer and audio encoder with Gemma 4, so running both together uses less memory than running them separately.
Key features:Text, code, images, video and audio in a single embedding space8K token context (4x EmbeddingGemma 1), enough for about 5.5 minutes of audio, 29 images or 58 video framesVectors can be shrunk from 768 to 512, 256 or 128 dimensions for up to 6x less storageCode retrieval improves from 68.76 to 78.68 on MTEB Code (+9.92 points)Google says it leads sub-1B multimodal embedding models on benchmarks like MTEB Code and MAEB (audio)Can also act as a zero-shot, on-device decision engine: Google's MediaPipe demo scores 500 options per turn in a chess game in under 100 msWorks with transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LMStudio, Qdrant and UnslothWho it's for: Developers building private, offline-capable search, RAG, media retrieval or local codebase search, and anyone who wants cross-modal search without a cloud dependency.
P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends
PH 用户
Love that EmbeddingGemma 2 maps text, images, audio and video into one shared embedding space so a single query can pull the right result from any of them.
PH 用户
Congrats on the launch! An open multimodal embedding model this light makes local retrieval much easier to ship. Which device sizes are you targeting for on-device use?
Problem: Searching by meaning usually needs a pipeline of separate models: a captioning model for images, speech-to-text for audio, a text embedder on top. That's slow, memory-hungry, and often sends your data to the cloud. For private, on-device apps, that's a hard trade-off.
Solution: One 740M-parameter model, built on the Gemma 4 architecture, that embeds all five types of content in the same space. A text query can find a photo, and a voice memo can find a video clip. Everything runs locally, so your data never has to leave the device.
What makes it different: It's modular. You can load only the encoders you need, as little as 270M parameters for text and code, plus optional vision (170M) and audio (300M) encoders. With quantization on a Pixel 11 Pro, it needs about 191MB of RAM for text only and about 567MB for the full multimodal model. It shares its tokenizer and audio encoder with Gemma 4, so running both together uses less memory than running them separately.
Key features:Text, code, images, video and audio in a single embedding space8K token context (4x EmbeddingGemma 1), enough for about 5.5 minutes of audio, 29 images or 58 video framesVectors can be shrunk from 768 to 512, 256 or 128 dimensions for up to 6x less storageCode retrieval improves from 68.76 to 78.68 on MTEB Code (+9.92 points)Google says it leads sub-1B multimodal embedding models on benchmarks like MTEB Code and MAEB (audio)Can also act as a zero-shot, on-device decision engine: Google's MediaPipe demo scores 500 options per turn in a chess game in under 100 msWorks with transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LMStudio, Qdrant and UnslothWho it's for: Developers building private, offline-capable search, RAG, media retrieval or local codebase search, and anyone who wants cross-modal search without a cloud dependency.
P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends