Google open-sources EmbeddingGemma 2: a 744M multimodal embedding model beating models 12x larger
dair_ai · x · 2026-10-28
Google DeepMind released EmbeddingGemma 2, an open 744M-parameter multimodal embedding model that maps text, images, audio, video, and mixed inputs into one shared vector space. Each modality has its own encoder feeding a shared backbone, with selective loading: text-only at 271M, vision-language at 439M, audio-language at 577M.
Key results:
- Beats specialist models and models 2–12x larger across seven benchmark suites (text, code, images, video, documents, audio)
- Scores 61.36 on MTEB Multilingual v2 and 68.46 on MTEB English v2
- int4 quantization costs only 0.4 points on MTEB Multilingual
It also supports reduced vision token budgets and truncatable embedding dimensions, making it a strong on-device option for agent memory and retrieval over mixed media.
Related event: Google Open-Sources EmbeddingGemma 2, a Compact Multimodal Embedding Model(2 posts)→
More from Models
- Researcher Drops Claude 5.5 for Biology Work Over Too Many Refusals — anshulkundaje · 2026-10-28
- Claude Opus 5.5 beats Opus 5 in 14 of 16 comparisons, leads physical agent Elo 1313-1289 — qinzytech · 2026-10-28
- Stanford prof: LLM reviewers catch mistakes scientists keep missing — anshulkundaje · 2026-10-28
- Correction: DeepSeek-V4.1-Flash-1017 scores only ~28% on FrontierCode 2.0 — scaling01 · 2026-10-28
- LiveBench update: Claude Fable 5.1 tops real-world leaderboard at 85.7, DeepSeek leads open source — bindureddy · 2026-10-28
- Open-source coding models look benchmaxxed; dev hopes FrontierCode 2.0 stays clean — zyxciss · 2026-10-28