EmbeddingGemma 2: An open, lightweight multimodal embedding model
Google says its new 740M-parameter embedding model can run multimodal retrieval locally across text, code, images, video, and audio.
EmbeddingGemma 2 maps mixed inputs into one embedding space and is released under Apache 2.0. Google says it can operate with modular encoders, from 270M parameters for text-only work to the full multimodal setup with vision and audio. With quantization, Google reports about 191MB active RAM for text-only weights and about 567MB for the full model on a Pixel 11 Pro. The post pitches it for on-device search, RAG, media retrieval, routing, and offline use.
Source: HN · Frontpage AI's note
EmbeddingGemma 2 maps mixed inputs into one embedding space and is released under Apache 2.0. Google says it can operate with modular encoders, from 270M parameters for text-only work to the full multimodal setup with vision and audio. With quantization, Google reports about 191MB active RAM for text-only weights and about 567MB for the full model on a Pixel 11 Pro. The post pitches it for on-device search, RAG, media retrieval, routing, and offline use.
Source: HN · Frontpage AI's note
score 7
Discussions
- hn · 256 points · 31 comments
- hn · 264 points · 31 comments
- hn · 273 points · 31 comments
- hn · 285 points · 31 comments
- hn · 290 points · 31 comments
- hn · 296 points · 31 comments
- hn · 301 points · 31 comments
- hn · 305 points · 31 comments
- hn · 313 points · 33 comments
- hn · 318 points · 34 comments
- hn · 328 points · 35 comments
- hn · 334 points · 35 comments
- hn · 342 points · 35 comments
- hn · 350 points · 35 comments
- hn · 356 points · 35 comments
- hn · 362 points · 36 comments
- hn · 370 points · 36 comments
- hn · 377 points · 37 comments
- hn · 383 points · 38 comments
- hn · 395 points · 41 comments