Repository navigation
model: EmbeddingGemma 2 (gemma-embedding2, 0.74B dense) #59
Description
A multimodal embedding model. It needs an embeddings endpoint, which bloomery lacks.
- Released 2026-09-14, Apache-2.0. The 300M EmbeddingGemma of 2025-07-17 is v1: https://huggingface.co/google/embeddinggemma-2
- 740M total: 270M text, 170M vision, 300M audio. One 768-d space; MRL 128/256/512/768; 8K context; mean pooling.
- 98,829 downloads, 1,736 likes; trending rank 1 for feature-extraction (read 2026-10-12).
- llama.cpp merged it on 2026-10-06: model: support embeddinggemma2 (text+vision+audio) ggml-org/llama.cpp#30054
GGUF: https://huggingface.co/ggml-org/embeddinggemma-2-GGUF (arch gemma-embedding2, non-causal; Q8_0 310 MB; mmproj Q8_0 555 MB); https://huggingface.co/unsloth/embeddinggemma-2-GGUF (UD-Q4_K_XL 176 MB).
Needs: /v1/embeddings and /embedding, mean pooling, a bidirectional encoder path, MRL cut and renorm, task prefixes. Plan-triage marks embeddings out of scope.
Open: GGUF metadata says context 262144, the card says 8,192. Side seat or skip?
Surveyed with an AI agent; facts link their sources.
Activity
midagedev commentedon Oct 12, 2026
Why we want this: bloomery is moving toward one server that holds several models at once. The unified tier (card, RAM and NVMe) would be one budget for the weights of all of them.
An embedding model is the smallest case. It stays on the card beside the chat model, and the plan takes its bytes out of the chat model's expert budget. So the answer to "side seat or skip" is a side seat.
The engine work is the list above: a bidirectional encoder path, mean pooling, the MRL cut, and
/v1/embeddings.midagedev commentedon Oct 20, 2026
Done: /v1/embeddings serves EmbeddingGemma 2 (crates/serve/src/embeddings.rs, PR #299).