Google Launches EmbeddingGemma 2, a 740M-Parameter On-Device Multimodal Embedding Model
Summary
Google launches EmbeddingGemma 2, a 740-million-parameter Apache 2.0 model that embeds text, code, images, video and audio in one on-device space, supporting 8K-token context and using as little as 191MB of active RAM on a Pixel 11 Pro when quantized.
Key Points
- Google launches EmbeddingGemma 2, a 740-million-parameter Apache 2.0 model that maps text, code, images, video and audio into one on-device embedding space.
- EmbeddingGemma 2 improves its MTEB Code score by 9.92 points, rising from 68.76 to 78.68, while Google says it leads sub-1B multimodal embedders on MTEB and MAEB benchmarks.
- The model supports an 8K-token context window and, with quantization on a Pixel 11 Pro, uses about 191MB of active RAM for text-only weights or 567MB for full multimodal inference.