The Reddit discussion points to Google’s release of EmbeddingGemma 2. Commenters describe it as a 740M model with native multimodality. Text, images, video and audio share a 768-dimensional embedding space. The model is described as using the Apache 2 license.

Comments point to multimodal indexing, media retrieval and RAG use cases. One user says the model runs on a Raspberry Pi. Another asks whether task-steered representations are a new feature. Users also mention Google endpoints, MRL, HQ-CLIP, EmbeddingGemma 1 and nomic-embed-text in existing workflows.

The most detailed evidence is a user-posted Claude comparison using MTG card art. It reports nDCG@20 for text search, nDCG@10 for image similarity and the reported model footprint. HQ-CLIP scored 0.715 and 0.834 at about 1.6 GB of fp32 RAM. EmbeddingGemma 1 in Q4 on captions scored 0.743 and 0.741 at about 0.3 GB. EmbeddingGemma 2 in BF16 on captions scored 0.775 and 0.797 at 0.56 GB. Its 4-bit caption version scored 0.771 and 0.801 at 0.18 GB.

The BF16 vision version scored 0.740 and 0.884 and used 0.56 GB plus a 0.98 GB mmproj. The 4-bit text version with a Q8 mmproj for vision scored 0.738 and 0.879 and used 0.18 GB plus 0.55 GB mmproj. The comparison found an approximate 0.004 nDCG drop after quantization and a 0.996 mean cosine similarity between the two vector sets.

The poster calculated a +0.061 nDCG@20 advantage over HQ-CLIP for caption search, with a 95% CI of +0.028 to +0.095. Vision reached +0.025 for text-to-image search, with a confidence interval that includes zero. Image-to-image similarity reached 0.884 versus 0.834 for HQ-CLIP. The caption-model image rows compare caption-to-caption results. These figures come from a community test. The excerpt provides no published model benchmark.